praat/external/espeak/readclause.cpp

/*
 * Copyright (C) 2005 to 2015 by Jonathan Duddington
 * email: jonsd@users.sourceforge.net
 * Copyright (C) 2015-2017 Reece H. Dunn
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, see: <http://www.gnu.org/licenses/>.
 */

#include <ctype.h>
#include <errno.h>
#include <locale.h>
#include <math.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <wchar.h>
#include <wctype.h>

#include "espeak_ng.h"
#include "encoding.h"
#include "ucd.h"

#include "error.h"
#include "speech.h"
#include "synthesize.h"
#include "translate.h"

#define N_XML_BUF   500

static const char *xmlbase = ""; // base URL from <speak>

static int namedata_ix = 0;
static int n_namedata = 0;
char *namedata = NULL;

static int ungot_char2 = 0;
espeak_ng_TEXT_DECODER *p_decoder = NULL;
static int ungot_char;
static const char *ungot_word = NULL;

static int ignore_text = 0; // set during <sub> ... </sub>  to ignore text which has been replaced by an alias
static int audio_text = 0; // set during <audio> ... </audio>
static int clear_skipping_text = 0; // next clause should clear the skipping_text flag
int count_characters = 0;
static int sayas_mode;
static int sayas_start;
static int ssml_ignore_l_angle = 0;

// stack for language and voice properties
// frame 0 is for the defaults, before any ssml tags.
typedef struct {
	int tag_type;
	int voice_variant_number;
	int voice_gender;
	int voice_age;
	char voice_name[40];
	char language[20];
} SSML_STACK;

#define N_SSML_STACK  20
static int n_ssml_stack;
static SSML_STACK ssml_stack[N_SSML_STACK];

static espeak_VOICE base_voice;
static char base_voice_variant_name[40] = { 0 };
static char current_voice_id[40] = { 0 };

#define N_PARAM_STACK  20
static int n_param_stack;
PARAM_STACK param_stack[N_PARAM_STACK];

static int speech_parameters[N_SPEECH_PARAM]; // current values, from param_stack
int saved_parameters[N_SPEECH_PARAM]; // Parameters saved on synthesis start

#define ESPEAKNG_CLAUSE_TYPE_PROPERTY_MASK 0xFFF0000000000000ull

int clause_type_from_codepoint(uint32_t c)
{
	ucd_category cat = ucd_lookup_category(c);
	ucd_property props = ucd_properties(c, cat);

	switch (props & ESPEAKNG_CLAUSE_TYPE_PROPERTY_MASK)
	{
	case ESPEAKNG_PROPERTY_FULL_STOP:
		return CLAUSE_PERIOD;
	case ESPEAKNG_PROPERTY_FULL_STOP | ESPEAKNG_PROPERTY_OPTIONAL_SPACE_AFTER:
		return CLAUSE_PERIOD | CLAUSE_OPTIONAL_SPACE_AFTER;
	case ESPEAKNG_PROPERTY_QUESTION_MARK:
		return CLAUSE_QUESTION;
	case ESPEAKNG_PROPERTY_QUESTION_MARK | ESPEAKNG_PROPERTY_OPTIONAL_SPACE_AFTER:
		return CLAUSE_QUESTION | CLAUSE_OPTIONAL_SPACE_AFTER;
	case ESPEAKNG_PROPERTY_QUESTION_MARK | ESPEAKNG_PROPERTY_PUNCTUATION_IN_WORD:
		return CLAUSE_QUESTION | CLAUSE_PUNCTUATION_IN_WORD;
	case ESPEAKNG_PROPERTY_EXCLAMATION_MARK:
		return CLAUSE_EXCLAMATION;
	case ESPEAKNG_PROPERTY_EXCLAMATION_MARK | ESPEAKNG_PROPERTY_OPTIONAL_SPACE_AFTER:
		return CLAUSE_EXCLAMATION | CLAUSE_OPTIONAL_SPACE_AFTER;
	case ESPEAKNG_PROPERTY_EXCLAMATION_MARK | ESPEAKNG_PROPERTY_PUNCTUATION_IN_WORD:
		return CLAUSE_EXCLAMATION | CLAUSE_PUNCTUATION_IN_WORD;
	case ESPEAKNG_PROPERTY_COMMA:
		return CLAUSE_COMMA;
	case ESPEAKNG_PROPERTY_COMMA | ESPEAKNG_PROPERTY_OPTIONAL_SPACE_AFTER:
		return CLAUSE_COMMA | CLAUSE_OPTIONAL_SPACE_AFTER;
	case ESPEAKNG_PROPERTY_COLON:
		return CLAUSE_COLON;
	case ESPEAKNG_PROPERTY_COLON | ESPEAKNG_PROPERTY_OPTIONAL_SPACE_AFTER:
		return CLAUSE_COLON | CLAUSE_OPTIONAL_SPACE_AFTER;
	case ESPEAKNG_PROPERTY_SEMI_COLON:
	case ESPEAKNG_PROPERTY_EXTENDED_DASH:
		return CLAUSE_SEMICOLON;
	case ESPEAKNG_PROPERTY_SEMI_COLON | ESPEAKNG_PROPERTY_OPTIONAL_SPACE_AFTER:
	case ESPEAKNG_PROPERTY_QUESTION_MARK | ESPEAKNG_PROPERTY_OPTIONAL_SPACE_AFTER | ESPEAKNG_PROPERTY_INVERTED_TERMINAL_PUNCTUATION:
	case ESPEAKNG_PROPERTY_EXCLAMATION_MARK | ESPEAKNG_PROPERTY_OPTIONAL_SPACE_AFTER | ESPEAKNG_PROPERTY_INVERTED_TERMINAL_PUNCTUATION:
		return CLAUSE_SEMICOLON | CLAUSE_OPTIONAL_SPACE_AFTER;
	case ESPEAKNG_PROPERTY_ELLIPSIS:
		return CLAUSE_SEMICOLON | CLAUSE_SPEAK_PUNCTUATION_NAME | CLAUSE_OPTIONAL_SPACE_AFTER;
	case ESPEAKNG_PROPERTY_PARAGRAPH_SEPARATOR:
		return CLAUSE_PARAGRAPH;
	}

	return CLAUSE_NONE;
}

const int param_defaults[N_SPEECH_PARAM] = {
	0,   // silence (internal use)
	175, // rate wpm
	100, // volume
	50,  // pitch
	50,  // range
	0,   // punctuation
	0,   // capital letters
	0,   // wordgap
	0,   // options
	0,   // intonation
	0,
	0,
	0,   // emphasis
	0,   // line length
	0,   // voice type
};

int towlower2(unsigned int c)
{
	// check for non-standard upper to lower case conversions
	if (c == 'I' && translator->langopts.dotless_i)
		return 0x131; // I -> ı

	return tolower(c);
}

static int IsRomanU(unsigned int c)
{
	if ((c == 'I') || (c == 'V') || (c == 'X') || (c == 'L'))
		return 1;
	return 0;
}

int Eof(void)
{
	if (ungot_char != 0)
		return 0;

	return text_decoder_eof(p_decoder);
}

static int GetC(void)
{
	int c1;

	if ((c1 = ungot_char) != 0) {
		ungot_char = 0;
		return c1;
	}

	count_characters++;
	return text_decoder_getc(p_decoder);
}

static void UngetC(int c)
{
	ungot_char = c;
}

const char *WordToString2(unsigned int word)
{
	// Convert a language mnemonic word into a string
	int ix;
	static char buf[5];
	char *p;

	p = buf;
	for (ix = 3; ix >= 0; ix--) {
		if ((*p = word >> (ix*8)) != 0)
			p++;
	}
	*p = 0;
	return buf;
}

static const char *LookupSpecial(Translator *tr, const char *string, char *text_out)
{
	unsigned int flags[2];
	char phonemes[55];
	char phonemes2[55];
	char *string1 = (char *)string;

	flags[0] = flags[1] = 0;
	if (LookupDictList(tr, &string1, phonemes, flags, 0, NULL)) {
		SetWordStress(tr, phonemes, flags, -1, 0);
		DecodePhonemes(phonemes, phonemes2);
		sprintf(text_out, "[\002%s]]", phonemes2);
		return text_out;
	}
	return NULL;
}

static const char *LookupCharName(Translator *tr, int c, int only)
{
	// Find the phoneme string (in ascii) to speak the name of character c
	// Used for punctuation characters and symbols

	int ix;
	unsigned int flags[2];
	char single_letter[24];
	char phonemes[60];
	char phonemes2[60];
	const char *lang_name = NULL;
	char *string;
	static char buf[60];

	buf[0] = 0;
	flags[0] = 0;
	flags[1] = 0;
	single_letter[0] = 0;
	single_letter[1] = '_';
	ix = utf8_out(c, &single_letter[2]);
	single_letter[2+ix] = 0;

	if (only) {
		string = &single_letter[2];
		LookupDictList(tr, &string, phonemes, flags, 0, NULL);
	} else {
		string = &single_letter[1];
		if (LookupDictList(tr, &string, phonemes, flags, 0, NULL) == 0) {
			// try _* then *
			string = &single_letter[2];
			if (LookupDictList(tr, &string, phonemes, flags, 0, NULL) == 0) {
				// now try the rules
				single_letter[1] = ' ';
				TranslateRules(tr, &single_letter[2], phonemes, sizeof(phonemes), NULL, 0, NULL);
			}
		}
	}

	if ((only == 0) && ((phonemes[0] == 0) || (phonemes[0] == phonSWITCH)) && (tr->translator_name != L('e', 'n'))) {
		// not found, try English
		SetTranslator2("en");
		string = &single_letter[1];
		single_letter[1] = '_';
		if (LookupDictList(translator2, &string, phonemes, flags, 0, NULL) == 0) {
			string = &single_letter[2];
			LookupDictList(translator2, &string, phonemes, flags, 0, NULL);
		}
		if (phonemes[0])
			lang_name = "en";
		else
			SelectPhonemeTable(voice->phoneme_tab_ix); // revert to original phoneme table
	}

	if (phonemes[0]) {
		if (lang_name) {
			SetWordStress(translator2, phonemes, flags, -1, 0);
			DecodePhonemes(phonemes, phonemes2);
			sprintf(buf, "[\002_^_%s %s _^_%s]]", "en", phonemes2, WordToString2(tr->translator_name));
			SelectPhonemeTable(voice->phoneme_tab_ix); // revert to original phoneme table
		} else {
			SetWordStress(tr, phonemes, flags, -1, 0);
			DecodePhonemes(phonemes, phonemes2);
			sprintf(buf, "[\002%s]] ", phonemes2);
		}
	} else if (only == 0)
		strcpy(buf, "[\002(X1)(X1)(X1)]]");

	return buf;
}

int Read4Bytes(FILE *f)
{
	// Read 4 bytes (least significant first) into a int32
	char p4 [4];
	fread (p4, 1, 4, f);
	return get_int32_le (p4);
}

static espeak_ng_STATUS LoadSoundFile(const char *fname, int index, espeak_ng_ERROR_CONTEXT *context)
{
	FILE *f;
	char *p;
	int *ip;
	int length;
	char fname_temp[100];
	char fname2[sizeof(path_home)+13+40];

	if (fname == NULL) {
		// filename is already in the table
		fname = soundicon_tab[index].filename;
	}

	if (fname == NULL)
		return static_cast<espeak_ng_STATUS> (EINVAL);

	if (fname[0] != '/') {
		// a relative path, look in espeak-ng-data/soundicons
		sprintf(fname2, "%s%csoundicons%c%s", path_home, PATHSEP, PATHSEP, fname);
		fname = fname2;
	}

	f = NULL;
	if ((f = fopen(fname, "rb")) != NULL) {
		int ix;
		int header[3];
		char command[sizeof(fname2)+sizeof(fname2)+40];

		if (fseek(f, 20, SEEK_SET) == -1) {
			int error = errno;
			fclose(f);
			return create_file_error_context(context, static_cast<espeak_ng_STATUS> (error), fname);
		}

		for (ix = 0; ix < 3; ix++)
			header[ix] = Read4Bytes(f);

		// if the sound file is not mono, 16 bit signed, at the correct sample rate, then convert it
		if ((header[0] != 0x10001) || (header[1] != samplerate) || (header[2] != samplerate*2)) {
			fclose(f);
			f = NULL;

#ifdef HAVE_MKSTEMP
			int fd_temp;
			strcpy(fname_temp, "/tmp/espeakXXXXXX");
			if ((fd_temp = mkstemp(fname_temp)) >= 0)
				close(fd_temp);
#else
			strcpy(fname_temp, tmpnam(NULL));
#endif

			sprintf(command, "sox \"%s\" -r %d -c1 -t wav %s\n", fname, samplerate, fname_temp);
			if (system(command) == 0)
				fname = fname_temp;
		}
	}

	if (f == NULL) {
		f = fopen(fname, "rb");
		if (f == NULL)
			return create_file_error_context(context, static_cast<espeak_ng_STATUS> (errno), fname);
	}

	length = GetFileLength(fname);
	if (length < 0) { // length == -errno
		fclose(f);
		return create_file_error_context(context, static_cast<espeak_ng_STATUS> (-length), fname);
	}
	if (fseek(f, 0, SEEK_SET) == -1) {
		int error = errno;
		fclose(f);
		return create_file_error_context(context, static_cast<espeak_ng_STATUS> (error), fname);
	}
	if ((p = (char *)realloc(soundicon_tab[index].data, length)) == NULL) {
		fclose(f);
// 		return static_cast<espeak_ng_STATUS> (ENOMEM);
	}
	if (fread(p, 1, length, f) != length) {
		int error = errno;
		fclose(f);
		remove(fname_temp);
		free(p);
		return create_file_error_context(context, static_cast<espeak_ng_STATUS> (error), fname);
	}
	fclose(f);
	remove(fname_temp);

	int i32 = get_int32_le (p + 40); // ip = (int *)(&p[40]);
	soundicon_tab[index].length = i32 / 2; // length in samples
	soundicon_tab[index].data = p;
	return ENS_OK;
}

static int LookupSoundicon(int c)
{
	// Find the sound icon number for a punctuation chatacter
	int ix;

	for (ix = N_SOUNDICON_SLOTS; ix < n_soundicon_tab; ix++) {
		if (soundicon_tab[ix].name == c) {
			if (soundicon_tab[ix].length == 0) {
				if (LoadSoundFile(NULL, ix, NULL) != ENS_OK)
					return -1; // sound file is not available
			}
			return ix;
		}
	}
	return -1;
}

static int LoadSoundFile2(const char *fname)
{
	// Load a sound file into one of the reserved slots in the sound icon table
	// (if it'snot already loaded)

	int ix;
	static int slot = -1;

	for (ix = 0; ix < n_soundicon_tab; ix++) {
		if (((soundicon_tab[ix].filename != NULL) && strcmp(fname, soundicon_tab[ix].filename) == 0))
			return ix; // already loaded
	}

	// load the file into the next slot
	slot++;
	if (slot >= N_SOUNDICON_SLOTS)
		slot = 0;

	if (LoadSoundFile(fname, slot, NULL) != ENS_OK)
		return -1;

	soundicon_tab[slot].filename = (char *)realloc(soundicon_tab[ix].filename, strlen(fname)+1);
	strcpy(soundicon_tab[slot].filename, fname);
	return slot;
}

static int AnnouncePunctuation(Translator *tr, int c1, int *c2_ptr, char *output, int *bufix, int end_clause)
{
	// announce punctuation names
	// c1:  the punctuation character
	// c2:  the following character

	int punct_count;
	const char *punctname = NULL;
	int soundicon;
	int attributes;
	int short_pause;
	int c2;
	int len;
	int bufix1;
	char buf[200];
	char buf2[80];
	char ph_buf[30];

	c2 = *c2_ptr;
	buf[0] = 0;

	if ((soundicon = LookupSoundicon(c1)) >= 0) {
		// add an embedded command to play the soundicon
		sprintf(buf, "\001%dI ", soundicon);
		UngetC(c2);
	} else {
		if ((c1 == '.') && (end_clause) && (c2 != '.')) {
			if (LookupSpecial(tr, "_.p", ph_buf))
				punctname = ph_buf; // use word for 'period' instead of 'dot'
		}
		if (punctname == NULL)
			punctname = LookupCharName(tr, c1, 0);

		if (punctname == NULL)
			return -1;

		if ((*bufix == 0) || (end_clause == 0) || (tr->langopts.param[LOPT_ANNOUNCE_PUNCT] & 2)) {
			punct_count = 1;
			while ((c2 == c1) && (c1 != '<')) { // don't eat extra '<', it can miss XML tags
				punct_count++;
				c2 = GetC();
			}
			*c2_ptr = c2;
			if (end_clause)
				UngetC(c2);

			if (punct_count == 1)
				sprintf(buf, " %s", punctname); // we need the space before punctname, to ensure it doesn't merge with the previous word  (eg.  "2.-a")
			else if (punct_count < 4) {
				buf[0] = 0;
				if (embedded_value[EMBED_S] < 300)
					sprintf(buf, "\001+10S"); // Speak punctuation name faster, unless we are already speaking fast.  It would upset Sonic SpeedUp

				while (punct_count-- > 0) {
					sprintf(buf2, " %s", punctname);
					strcat(buf, buf2);
				}

				if (embedded_value[EMBED_S] < 300) {
					sprintf(buf2, " \001-10S");
					strcat(buf, buf2);
				}
			} else
				sprintf(buf, " %s %d %s",
				        punctname, punct_count, punctname);
		} else {
			// end the clause now and pick up the punctuation next time
			UngetC(c2);
			if (option_ssml) {
				if ((c1 == '<') || (c1 == '&'))
					ssml_ignore_l_angle = c1; // this was &lt; which was converted to <, don't pick it up again as <
			}
			ungot_char2 = c1;
			buf[0] = ' ';
			buf[1] = 0;
		}
	}

	bufix1 = *bufix;
	len = strlen(buf);
	strcpy(&output[*bufix], buf);
	*bufix += len;

	if (end_clause == 0)
		return -1;

	if (c1 == '-')
		return CLAUSE_NONE; // no pause

	attributes = clause_type_from_codepoint(c1);

	short_pause = CLAUSE_SHORTFALL;
	if ((attributes & CLAUSE_INTONATION_TYPE) == 0x1000)
		short_pause = CLAUSE_SHORTCOMMA;

	if ((bufix1 > 0) && !(tr->langopts.param[LOPT_ANNOUNCE_PUNCT] & 2)) {
		if ((attributes & ~CLAUSE_OPTIONAL_SPACE_AFTER) == CLAUSE_SEMICOLON)
			return CLAUSE_SHORTFALL;
		return short_pause;
	}

	if (attributes & CLAUSE_TYPE_SENTENCE)
		return attributes;

	return short_pause;
}

#define SSML_SPEAK     1
#define SSML_VOICE     2
#define SSML_PROSODY   3
#define SSML_SAYAS     4
#define SSML_MARK      5
#define SSML_SENTENCE  6
#define SSML_PARAGRAPH 7
#define SSML_PHONEME   8
#define SSML_SUB       9
#define SSML_STYLE    10
#define SSML_AUDIO    11
#define SSML_EMPHASIS 12
#define SSML_BREAK    13
#define SSML_IGNORE_TEXT 14
#define HTML_BREAK    15
#define HTML_NOSPACE  16   // don't insert a space for this element, so it doesn't break a word
#define SSML_CLOSE    0x20 // for a closing tag, OR this with the tag type

// these tags have no effect if they are self-closing, eg. <voice />
static char ignore_if_self_closing[] = { 0, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 1, 0, 1, 0, 0, 0, 0 };

static MNEM_TAB ssmltags[] = {
	{ "speak",     SSML_SPEAK },
	{ "voice",     SSML_VOICE },
	{ "prosody",   SSML_PROSODY },
	{ "say-as",    SSML_SAYAS },
	{ "mark",      SSML_MARK },
	{ "s",         SSML_SENTENCE },
	{ "p",         SSML_PARAGRAPH },
	{ "phoneme",   SSML_PHONEME },
	{ "sub",       SSML_SUB },
	{ "tts:style", SSML_STYLE },
	{ "audio",     SSML_AUDIO },
	{ "emphasis",  SSML_EMPHASIS },
	{ "break",     SSML_BREAK },
	{ "metadata",  SSML_IGNORE_TEXT },

	{ "br",     HTML_BREAK },
	{ "li",     HTML_BREAK },
	{ "dd",     HTML_BREAK },
	{ "img",    HTML_BREAK },
	{ "td",     HTML_BREAK },
	{ "h1",     SSML_PARAGRAPH },
	{ "h2",     SSML_PARAGRAPH },
	{ "h3",     SSML_PARAGRAPH },
	{ "h4",     SSML_PARAGRAPH },
	{ "hr",     SSML_PARAGRAPH },
	{ "script", SSML_IGNORE_TEXT },
	{ "style",  SSML_IGNORE_TEXT },
	{ "font",   HTML_NOSPACE },
	{ "b",      HTML_NOSPACE },
	{ "i",      HTML_NOSPACE },
	{ "strong", HTML_NOSPACE },
	{ "em",     HTML_NOSPACE },
	{ "code",   HTML_NOSPACE },

	{ NULL, 0 }
};

static const char *VoiceFromStack()
{
	// Use the voice properties from the SSML stack to choose a voice, and switch
	// to that voice if it's not the current voice

	int ix;
	const char *p;
	SSML_STACK *sp;
	const char *v_id;
	int voice_name_specified;
	int voice_found;
	espeak_VOICE voice_select;
	static char voice_name[40];
	char language[40];
	char buf[80];

	strcpy(voice_name, ssml_stack[0].voice_name);
	strcpy(language, ssml_stack[0].language);
	voice_select.age = ssml_stack[0].voice_age;
	voice_select.gender = ssml_stack[0].voice_gender;
	voice_select.variant = ssml_stack[0].voice_variant_number;
	voice_select.identifier = NULL;

	for (ix = 0; ix < n_ssml_stack; ix++) {
		sp = &ssml_stack[ix];
		voice_name_specified = 0;

		if ((sp->voice_name[0] != 0) && (SelectVoiceByName(NULL, sp->voice_name) != NULL)) {
			voice_name_specified = 1;
			strcpy(voice_name, sp->voice_name);
			language[0] = 0;
			voice_select.gender = ENGENDER_UNKNOWN;
			voice_select.age = 0;
			voice_select.variant = 0;
		}
		if (sp->language[0] != 0) {
			strcpy(language, sp->language);

			// is this language provided by the base voice?
			p = base_voice.languages;
			while (*p++ != 0) {
				if (strcmp(p, language) == 0) {
					// yes, change the language to the main language of the base voice
					strcpy(language, &base_voice.languages[1]);
					break;
				}
				p += (strlen(p) + 1);
			}

			if (voice_name_specified == 0)
				voice_name[0] = 0; // forget a previous voice name if a language is specified
		}
		if (sp->voice_gender != ENGENDER_UNKNOWN)
			voice_select.gender = sp->voice_gender;

		if (sp->voice_age != 0)
			voice_select.age = sp->voice_age;
		if (sp->voice_variant_number != 0)
			voice_select.variant = sp->voice_variant_number;
	}

	voice_select.name = voice_name;
	voice_select.languages = language;
	v_id = SelectVoice(&voice_select, &voice_found);
	if (v_id == NULL)
		return "default";

	if ((strchr(v_id, '+') == NULL) && ((voice_select.gender == ENGENDER_UNKNOWN) || (voice_select.gender == base_voice.gender)) && (base_voice_variant_name[0] != 0)) {
		// a voice variant has not been selected, use the original voice variant
		sprintf(buf, "%s+%s", v_id, base_voice_variant_name);
		strncpy0(voice_name, buf, sizeof(voice_name));
		return voice_name;
	}
	return v_id;
}

static void ProcessParamStack(char *outbuf, int *outix)
{
	// Set the speech parameters from the parameter stack
	int param;
	int ix;
	int value;
	char buf[20];
	int new_parameters[N_SPEECH_PARAM];
	static char cmd_letter[N_SPEECH_PARAM] = { 0, 'S', 'A', 'P', 'R', 0, 'C', 0, 0, 0, 0, 0, 'F' }; // embedded command letters

	for (param = 0; param < N_SPEECH_PARAM; param++)
		new_parameters[param] = -1;

	for (ix = 0; ix < n_param_stack; ix++) {
		for (param = 0; param < N_SPEECH_PARAM; param++) {
			if (param_stack[ix].parameter[param] >= 0)
				new_parameters[param] = param_stack[ix].parameter[param];
		}
	}

	for (param = 0; param < N_SPEECH_PARAM; param++) {
		if ((value = new_parameters[param]) != speech_parameters[param]) {
			buf[0] = 0;

			switch (param)
			{
			case espeakPUNCTUATION:
				option_punctuation = value-1;
				break;
			case espeakCAPITALS:
				option_capitals = value;
				break;
			case espeakRATE:
			case espeakVOLUME:
			case espeakPITCH:
			case espeakRANGE:
			case espeakEMPHASIS:
				sprintf(buf, "%c%d%c", CTRL_EMBEDDED, value, cmd_letter[param]);
				break;
			}

			speech_parameters[param] = new_parameters[param];
			strcpy(&outbuf[*outix], buf);
			*outix += strlen(buf);
		}
	}
}

static PARAM_STACK *PushParamStack(int tag_type)
{
	int ix;
	PARAM_STACK *sp;

	sp = &param_stack[n_param_stack];
	if (n_param_stack < (N_PARAM_STACK-1))
		n_param_stack++;

	sp->type = tag_type;
	for (ix = 0; ix < N_SPEECH_PARAM; ix++)
		sp->parameter[ix] = -1;
	return sp;
}

static void PopParamStack(int tag_type, char *outbuf, int *outix)
{
	// unwind the stack up to and including the previous tag of this type
	int ix;
	int top = 0;

	if (tag_type >= SSML_CLOSE)
		tag_type -= SSML_CLOSE;

	for (ix = 0; ix < n_param_stack; ix++) {
		if (param_stack[ix].type == tag_type)
			top = ix;
	}
	if (top > 0)
		n_param_stack = top;
	ProcessParamStack(outbuf, outix);
}

static wchar_t *GetSsmlAttribute(wchar_t *pw, const char *name)
{
	// Gets the value string for an attribute.
	// Returns NULL if the attribute is not present

	int ix;
	static wchar_t empty[1] = { 0 };

	while (*pw != 0) {
		if (iswspace(pw[-1])) {
			ix = 0;
			while (*pw == name[ix]) {
				pw++;
				ix++;
			}
			if (name[ix] == 0) {
				// found the attribute, now get the value
				while (iswspace(*pw)) pw++;
				if (*pw == '=') pw++;
				while (iswspace(*pw)) pw++;
				if ((*pw == '"') || (*pw == '\'')) // allow single-quotes ?
					return pw+1;
				else
					return empty;
			}
		}
		pw++;
	}
	return NULL;
}

static int attrcmp(const wchar_t *string1, const char *string2)
{
	int ix;

	if (string1 == NULL)
		return 1;

	for (ix = 0; (string1[ix] == string2[ix]) && (string1[ix] != 0); ix++)
		;
	if (((string1[ix] == '"') || (string1[ix] == '\'')) && (string2[ix] == 0))
		return 0;
	return 1;
}

static int attrlookup(const wchar_t *string1, const MNEM_TAB *mtab)
{
	int ix;

	for (ix = 0; mtab[ix].mnem != NULL; ix++) {
		if (attrcmp(string1, mtab[ix].mnem) == 0)
			return mtab[ix].value;
	}
	return mtab[ix].value;
}

static int attrnumber(const wchar_t *pw, int default_value, int type)
{
	int value = 0;

	if ((pw == NULL) || !IsDigit09(*pw))
		return default_value;

	while (IsDigit09(*pw))
		value = value*10 + *pw++ - '0';
	if ((type == 1) && (tolower(*pw) == 's')) {
		// time: seconds rather than ms
		value *= 1000;
	}
	return value;
}

static int attrcopy_utf8(char *buf, const wchar_t *pw, int len)
{
	// Convert attribute string into utf8, write to buf, and return its utf8 length
	unsigned int c;
	int ix = 0;
	int n;
	int prev_c = 0;

	if (pw != NULL) {
		while ((ix < (len-4)) && ((c = *pw++) != 0)) {
			if ((c == '"') && (prev_c != '\\'))
				break; // " indicates end of attribute, unless preceded by backstroke
			n = utf8_out(c, &buf[ix]);
			ix += n;
			prev_c = c;
		}
	}
	buf[ix] = 0;
	return ix;
}

static int attr_prosody_value(int param_type, const wchar_t *pw, int *value_out)
{
	int sign = 0;
	wchar_t *tail;
	double value;

	while (iswspace(*pw)) pw++;
	if (*pw == '+') {
		pw++;
		sign = 1;
	}
	if (*pw == '-') {
		pw++;
		sign = -1;
	}
	value = (double)wcstod(pw, &tail);
	if (tail == pw) {
		// failed to find a number, return 100%
		*value_out = 100;
		return 2;
	}

	if (*tail == '%') {
		if (sign != 0)
			value = 100 + (sign * value);
		*value_out = (int)value;
		return 2; // percentage
	}

	if ((tail[0] == 's') && (tail[1] == 't')) {
		double x;
		// convert from semitones to a  frequency percentage
		x = pow((double)2.0, (double)((value*sign)/12)) * 100;
		*value_out = (int)x;
		return 2; // percentage
	}

	if (param_type == espeakRATE) {
		if (sign == 0)
			*value_out = (int)(value * 100);
		else
			*value_out = 100 + (int)(sign * value * 100);
		return 2; // percentage
	}

	*value_out = (int)value;
	return sign;   // -1, 0, or 1
}

static int AddNameData(const char *name, int wide)
{
	// Add the name to the namedata and return its position
	// (Used by the Windows SAPI wrapper)

	int ix;
	int len;
	void *vp;

	if (wide) {
		len = (wcslen((const wchar_t *)name)+1)*sizeof(wchar_t);
		n_namedata = (n_namedata + sizeof(wchar_t) - 1) % sizeof(wchar_t);  // round to wchar_t boundary
	} else
		len = strlen(name)+1;

	if (namedata_ix+len >= n_namedata) {
		// allocate more space for marker names
		if ((vp = realloc(namedata, namedata_ix+len + 1000)) == NULL)
			return -1;  // failed to allocate, original data is unchanged but ignore this new name
		// !!! Bug?? If the allocated data shifts position, then pointers given to user application will be invalid

		namedata = (char *)vp;
		n_namedata = namedata_ix+len + 1000;
	}
	memcpy(&namedata[ix = namedata_ix], name, len);
	namedata_ix += len;
	return ix;
}

void SetVoiceStack(espeak_VOICE *v, const char *variant_name)
{
	SSML_STACK *sp;
	sp = &ssml_stack[0];

	if (v == NULL) {
		memset(sp, 0, sizeof(ssml_stack[0]));
		return;
	}
	if (v->languages != NULL)
		strcpy(sp->language, v->languages);
	if (v->name != NULL)
		strncpy0(sp->voice_name, v->name, sizeof(sp->voice_name));
	sp->voice_variant_number = v->variant;
	sp->voice_age = v->age;
	sp->voice_gender = v->gender;

	if (memcmp(variant_name, "!v", 2) == 0)
		variant_name += 3; // strip variant directory name, !v plus PATHSEP
	strncpy0(base_voice_variant_name, variant_name, sizeof(base_voice_variant_name));
	memcpy(&base_voice, &current_voice_selected, sizeof(base_voice));
}

static int GetVoiceAttributes(wchar_t *pw, int tag_type)
{
	// Determines whether voice attribute are specified in this tag, and if so, whether this means
	// a voice change.
	// If it's a closing tag, delete the top frame of the stack and determine whether this implies
	// a voice change.
	// Returns  CLAUSE_TYPE_VOICE_CHANGE if there is a voice change

	wchar_t *lang;
	wchar_t *gender;
	wchar_t *name;
	wchar_t *age;
	wchar_t *variant;
	int value;
	const char *new_voice_id;
	SSML_STACK *ssml_sp;

	static const MNEM_TAB mnem_gender[] = {
		{ "male", ENGENDER_MALE },
		{ "female", ENGENDER_FEMALE },
		{ "neutral", ENGENDER_NEUTRAL },
		{ NULL, ENGENDER_UNKNOWN }
	};

	if (tag_type & SSML_CLOSE) {
		// delete a stack frame
		if (n_ssml_stack > 1)
			n_ssml_stack--;
	} else {
		// add a stack frame if any voice details are specified
		lang = GetSsmlAttribute(pw, "xml:lang");

		if (tag_type != SSML_VOICE) {
			// only expect an xml:lang attribute
			name = NULL;
			variant = NULL;
			age = NULL;
			gender = NULL;
		} else {
			name = GetSsmlAttribute(pw, "name");
			variant = GetSsmlAttribute(pw, "variant");
			age = GetSsmlAttribute(pw, "age");
			gender = GetSsmlAttribute(pw, "gender");
		}

		if ((tag_type != SSML_VOICE) && (lang == NULL))
			return 0; // <s> or <p> without language spec, nothing to do

		ssml_sp = &ssml_stack[n_ssml_stack++];

		attrcopy_utf8(ssml_sp->language, lang, sizeof(ssml_sp->language));
		attrcopy_utf8(ssml_sp->voice_name, name, sizeof(ssml_sp->voice_name));
		if ((value = attrnumber(variant, 1, 0)) > 0)
			value--; // variant='0' and variant='1' the same
		ssml_sp->voice_variant_number = value;
		ssml_sp->voice_age = attrnumber(age, 0, 0);
		ssml_sp->voice_gender = attrlookup(gender, mnem_gender);
		ssml_sp->tag_type = tag_type;
	}

	new_voice_id = VoiceFromStack();
	if (strcmp(new_voice_id, current_voice_id) != 0) {
		// add an embedded command to change the voice
		strcpy(current_voice_id, new_voice_id);
		return CLAUSE_TYPE_VOICE_CHANGE;
	}

	return 0;
}

static void SetProsodyParameter(int param_type, wchar_t *attr1, PARAM_STACK *sp)
{
	int value;
	int sign;

	static const MNEM_TAB mnem_volume[] = {
		{ "default", 100 },
		{ "silent",    0 },
		{ "x-soft",   30 },
		{ "soft",     65 },
		{ "medium",  100 },
		{ "loud",    150 },
		{ "x-loud",  230 },
		{ NULL,       -1 }
	};

	static const MNEM_TAB mnem_rate[] = {
		{ "default", 100 },
		{ "x-slow",   60 },
		{ "slow",     80 },
		{ "medium",  100 },
		{ "fast",    125 },
		{ "x-fast",  160 },
		{ NULL,       -1 }
	};

	static const MNEM_TAB mnem_pitch[] = {
		{ "default", 100 },
		{ "x-low",    70 },
		{ "low",      85 },
		{ "medium",  100 },
		{ "high",    110 },
		{ "x-high",  120 },
		{ NULL,       -1 }
	};

	static const MNEM_TAB mnem_range[] = {
		{ "default", 100 },
		{ "x-low",    20 },
		{ "low",      50 },
		{ "medium",  100 },
		{ "high",    140 },
		{ "x-high",  180 },
		{ NULL,       -1 }
	};

	static const MNEM_TAB *mnem_tabs[5] = {
		NULL, mnem_rate, mnem_volume, mnem_pitch, mnem_range
	};

	if ((value = attrlookup(attr1, mnem_tabs[param_type])) >= 0) {
		// mnemonic specifies a value as a percentage of the base pitch/range/rate/volume
		sp->parameter[param_type] = (param_stack[0].parameter[param_type] * value)/100;
	} else {
		sign = attr_prosody_value(param_type, attr1, &value);

		if (sign == 0)
			sp->parameter[param_type] = value; // absolute value in Hz
		else if (sign == 2) {
			// change specified as percentage or in semitones
			sp->parameter[param_type] = (speech_parameters[param_type] * value)/100;
		} else {
			// change specified as plus or minus Hz
			sp->parameter[param_type] = speech_parameters[param_type] + (value*sign);
		}
	}
}

static int ReplaceKeyName(char *outbuf, int index, int *outix)
{
	// Replace some key-names by single characters, so they can be pronounced in different languages
	static MNEM_TAB keynames[] = {
		{ "space ",        0xe020 },
		{ "tab ",          0xe009 },
		{ "underscore ",   0xe05f },
		{ "double-quote ", '"' },
		{ NULL,            0 }
	};

	int ix;
	int letter;
	char *p;

	p = &outbuf[index];

	if ((letter = LookupMnem(keynames, p)) != 0) {
		ix = utf8_out(letter, p);
		*outix = index + ix;
		return letter;
	}
	return 0;
}

static int ProcessSsmlTag(wchar_t *xml_buf, char *outbuf, int *outix, int n_outbuf, int self_closing)
{
	// xml_buf is the tag and attributes with a zero terminator in place of the original '>'
	// returns a clause terminator value.

	unsigned int ix;
	int index;
	int c;
	int tag_type;
	int value;
	int value2;
	int value3;
	int voice_change_flag;
	wchar_t *px;
	wchar_t *attr1;
	wchar_t *attr2;
	wchar_t *attr3;
	int terminator;
	char *uri;
	int param_type;
	char tag_name[40];
	char buf[80];
	PARAM_STACK *sp;
	SSML_STACK *ssml_sp;

	static const MNEM_TAB mnem_phoneme_alphabet[] = {
		{ "espeak", 1 },
		{ NULL,    -1 }
	};

	static const MNEM_TAB mnem_punct[] = {
		{ "none", 1 },
		{ "all",  2 },
		{ "some", 3 },
		{ NULL,  -1 }
	};

	static const MNEM_TAB mnem_capitals[] = {
		{ "no",        0 },
		{ "icon",      1 },
		{ "spelling",  2 },
		{ "pitch",    20 },  // this is the amount by which to raise the pitch
		{ NULL,       -1 }
	};

	static const MNEM_TAB mnem_interpret_as[] = {
		{ "characters", SAYAS_CHARS },
		{ "tts:char",   SAYAS_SINGLE_CHARS },
		{ "tts:key",    SAYAS_KEY },
		{ "tts:digits", SAYAS_DIGITS },
		{ "telephone",  SAYAS_DIGITS1 },
		{ NULL,         -1 }
	};

	static const MNEM_TAB mnem_sayas_format[] = {
		{ "glyphs", 1 },
		{ NULL,    -1 }
	};

	static const MNEM_TAB mnem_break[] = {
		{ "none",     0 },
		{ "x-weak",   1 },
		{ "weak",     2 },
		{ "medium",   3 },
		{ "strong",   4 },
		{ "x-strong", 5 },
		{ NULL,      -1 }
	};

	static const MNEM_TAB mnem_emphasis[] = {
		{ "none",     1 },
		{ "reduced",  2 },
		{ "moderate", 3 },
		{ "strong",   4 },
		{ "x-strong", 5 },
		{ NULL,      -1 }
	};

	static const char *prosody_attr[5] = {
		NULL, "rate", "volume", "pitch", "range"
	};

	for (ix = 0; ix < (sizeof(tag_name)-1); ix++) {
		if (((c = xml_buf[ix]) == 0) || iswspace(c))
			break;
		tag_name[ix] = tolower((char)c);
	}
	tag_name[ix] = 0;

	px = &xml_buf[ix]; // the tag's attributes

	if (tag_name[0] == '/') {
		// closing tag
		if ((tag_type = LookupMnem(ssmltags, &tag_name[1])) != HTML_NOSPACE)
			outbuf[(*outix)++] = ' ';
		tag_type += SSML_CLOSE;
	} else {
		if ((tag_type = LookupMnem(ssmltags, tag_name)) != HTML_NOSPACE) {
			// separate SSML tags from the previous word (but not HMTL tags such as <b> <font> which can occur inside a word)
			outbuf[(*outix)++] = ' ';
		}

		if (self_closing && ignore_if_self_closing[tag_type])
			return 0;
	}

	voice_change_flag = 0;
	ssml_sp = &ssml_stack[n_ssml_stack-1];

	switch (tag_type)
	{
	case SSML_STYLE:
		sp = PushParamStack(tag_type);
		attr1 = GetSsmlAttribute(px, "field");
		attr2 = GetSsmlAttribute(px, "mode");


		if (attrcmp(attr1, "punctuation") == 0) {
			value = attrlookup(attr2, mnem_punct);
			sp->parameter[espeakPUNCTUATION] = value;
		} else if (attrcmp(attr1, "capital_letters") == 0) {
			value = attrlookup(attr2, mnem_capitals);
			sp->parameter[espeakCAPITALS] = value;
		}
		ProcessParamStack(outbuf, outix);
		break;
	case SSML_PROSODY:
		sp = PushParamStack(tag_type);

		// look for attributes:  rate, volume, pitch, range
		for (param_type = espeakRATE; param_type <= espeakRANGE; param_type++) {
			if ((attr1 = GetSsmlAttribute(px, prosody_attr[param_type])) != NULL)
				SetProsodyParameter(param_type, attr1, sp);
		}

		ProcessParamStack(outbuf, outix);
		break;
	case SSML_EMPHASIS:
		sp = PushParamStack(tag_type);
		value = 3; // default is "moderate"
		if ((attr1 = GetSsmlAttribute(px, "level")) != NULL)
			value = attrlookup(attr1, mnem_emphasis);

		if (translator->langopts.tone_language == 1) {
			static unsigned char emphasis_to_pitch_range[] = { 50, 50, 40, 70, 90, 100 };
			static unsigned char emphasis_to_volume[] = { 100, 100, 70, 110, 135, 150 };
			// tone language (eg.Chinese) do emphasis by increasing the pitch range.
			sp->parameter[espeakRANGE] = emphasis_to_pitch_range[value];
			sp->parameter[espeakVOLUME] = emphasis_to_volume[value];
		} else {
			static unsigned char emphasis_to_volume2[] = { 100, 100, 75, 100, 120, 150 };
			sp->parameter[espeakVOLUME] = emphasis_to_volume2[value];
			sp->parameter[espeakEMPHASIS] = value;
		}
		ProcessParamStack(outbuf, outix);
		break;
	case SSML_STYLE + SSML_CLOSE:
	case SSML_PROSODY + SSML_CLOSE:
	case SSML_EMPHASIS + SSML_CLOSE:
		PopParamStack(tag_type, outbuf, outix);
		break;
	case SSML_PHONEME:
		attr1 = GetSsmlAttribute(px, "alphabet");
		attr2 = GetSsmlAttribute(px, "ph");
		value = attrlookup(attr1, mnem_phoneme_alphabet);
		if (value == 1) { // alphabet="espeak"
			outbuf[(*outix)++] = '[';
			outbuf[(*outix)++] = '[';
			*outix += attrcopy_utf8(&outbuf[*outix], attr2, n_outbuf-*outix);
			outbuf[(*outix)++] = ']';
			outbuf[(*outix)++] = ']';
		}
		break;
	case SSML_SAYAS:
		attr1 = GetSsmlAttribute(px, "interpret-as");
		attr2 = GetSsmlAttribute(px, "format");
		attr3 = GetSsmlAttribute(px, "detail");
		value = attrlookup(attr1, mnem_interpret_as);
		value2 = attrlookup(attr2, mnem_sayas_format);
		if (value2 == 1)
			value = SAYAS_GLYPHS;

		value3 = attrnumber(attr3, 0, 0);

		if (value == SAYAS_DIGITS) {
			if (value3 <= 1)
				value = SAYAS_DIGITS1;
			else
				value = SAYAS_DIGITS + value3;
		}

		sprintf(buf, "%c%dY", CTRL_EMBEDDED, value);
		strcpy(&outbuf[*outix], buf);
		*outix += strlen(buf);

		sayas_start = *outix;
		sayas_mode = value; // punctuation doesn't end clause during SAY-AS
		break;
	case SSML_SAYAS + SSML_CLOSE:
		if (sayas_mode == SAYAS_KEY) {
			outbuf[*outix] = 0;
			ReplaceKeyName(outbuf, sayas_start, outix);
		}

		outbuf[(*outix)++] = CTRL_EMBEDDED;
		outbuf[(*outix)++] = 'Y';
		sayas_mode = 0;
		break;
	case SSML_SUB:
		if ((attr1 = GetSsmlAttribute(px, "alias")) != NULL) {
			// use the alias  rather than the text
			ignore_text = 1;
			*outix += attrcopy_utf8(&outbuf[*outix], attr1, n_outbuf-*outix);
		}
		break;
	case SSML_IGNORE_TEXT:
		ignore_text = 1;
		break;
	case SSML_SUB + SSML_CLOSE:
	case SSML_IGNORE_TEXT + SSML_CLOSE:
		ignore_text = 0;
		break;
	case SSML_MARK:
		if ((attr1 = GetSsmlAttribute(px, "name")) != NULL) {
			// add name to circular buffer of marker names
			attrcopy_utf8(buf, attr1, sizeof(buf));

			if (strcmp(skip_marker, buf) == 0) {
				// This is the marker we are waiting for before starting to speak
				clear_skipping_text = 1;
				skip_marker[0] = 0;
				return CLAUSE_NONE;
			}

			if ((index = AddNameData(buf, 0)) >= 0) {
				sprintf(buf, "%c%dM", CTRL_EMBEDDED, index);
				strcpy(&outbuf[*outix], buf);
				*outix += strlen(buf);
			}
		}
		break;
	case SSML_AUDIO:
		sp = PushParamStack(tag_type);

		if ((attr1 = GetSsmlAttribute(px, "src")) != NULL) {
			char fname[256];
			attrcopy_utf8(buf, attr1, sizeof(buf));

			if (uri_callback == NULL) {
				if ((xmlbase != NULL) && (buf[0] != '/')) {
					sprintf(fname, "%s/%s", xmlbase, buf);
					index = LoadSoundFile2(fname);
				} else
					index = LoadSoundFile2(buf);
				if (index >= 0) {
					sprintf(buf, "%c%dI", CTRL_EMBEDDED, index);
					strcpy(&outbuf[*outix], buf);
					*outix += strlen(buf);
					sp->parameter[espeakSILENCE] = 1;
				}
			} else {
				if ((index = AddNameData(buf, 0)) >= 0) {
					uri = &namedata[index];
					if (uri_callback(1, uri, xmlbase) == 0) {
						sprintf(buf, "%c%dU", CTRL_EMBEDDED, index);
						strcpy(&outbuf[*outix], buf);
						*outix += strlen(buf);
						sp->parameter[espeakSILENCE] = 1;
					}
				}
			}
		}
		ProcessParamStack(outbuf, outix);

		if (self_closing)
			PopParamStack(tag_type, outbuf, outix);
		else
			audio_text = 1;
		return CLAUSE_NONE;
	case SSML_AUDIO + SSML_CLOSE:
		PopParamStack(tag_type, outbuf, outix);
		audio_text = 0;
		return CLAUSE_NONE;
	case SSML_BREAK:
		value = 21;
		terminator = CLAUSE_NONE;

		if ((attr1 = GetSsmlAttribute(px, "strength")) != NULL) {
			static int break_value[6] = { 0, 7, 14, 21, 40, 80 }; // *10mS
			value = attrlookup(attr1, mnem_break);
			if (value < 3) {
				// adjust prepause on the following word
				sprintf(&outbuf[*outix], "%c%dB", CTRL_EMBEDDED, value);
				*outix += 3;
				terminator = 0;
			}
			value = break_value[value];
		}
		if ((attr2 = GetSsmlAttribute(px, "time")) != NULL) {
			value2 = attrnumber(attr2, 0, 1);   // pause in mS

			// compensate for speaking speed to keep constant pause length, see function PauseLength()
			// 'value' here is x 10mS
			value = (value2 * 256) / (speed.clause_pause_factor * 10);
			if (value < 200)
				value = (value2 * 256) / (speed.pause_factor * 10);

			if (terminator == 0)
				terminator = CLAUSE_NONE;
		}
		if (terminator) {
			if (value > 0xfff) {
				// scale down the value and set a scaling indicator bit
				value = value / 32;
				if (value > 0xfff)
					value = 0xfff;
				terminator |= CLAUSE_PAUSE_LONG;
			}
			return terminator + value;
		}
		break;
	case SSML_SPEAK:
		if ((attr1 = GetSsmlAttribute(px, "xml:base")) != NULL) {
			attrcopy_utf8(buf, attr1, sizeof(buf));
			if ((index = AddNameData(buf, 0)) >= 0)
				xmlbase = &namedata[index];
		}
		if (GetVoiceAttributes(px, tag_type) == 0)
			return 0; // no voice change
		return CLAUSE_VOICE;
	case SSML_VOICE:
		if (GetVoiceAttributes(px, tag_type) == 0)
			return 0; // no voice change
		return CLAUSE_VOICE;
	case SSML_SPEAK + SSML_CLOSE:
		// unwind stack until the previous <voice> or <speak> tag
		while ((n_ssml_stack > 1) && (ssml_stack[n_ssml_stack-1].tag_type != SSML_SPEAK))
			n_ssml_stack--;
		return CLAUSE_PERIOD + GetVoiceAttributes(px, tag_type);
	case SSML_VOICE + SSML_CLOSE:
		// unwind stack until the previous <voice> or <speak> tag
		while ((n_ssml_stack > 1) && (ssml_stack[n_ssml_stack-1].tag_type != SSML_VOICE))
			n_ssml_stack--;

		terminator = 0; // ??  Sentence intonation, but no pause ??
		return terminator + GetVoiceAttributes(px, tag_type);
	case HTML_BREAK:
	case HTML_BREAK + SSML_CLOSE:
		return CLAUSE_COLON;
	case SSML_SENTENCE:
		if (ssml_sp->tag_type == SSML_SENTENCE) {
			// new sentence implies end-of-sentence
			voice_change_flag = GetVoiceAttributes(px, SSML_SENTENCE+SSML_CLOSE);
		}
		voice_change_flag |= GetVoiceAttributes(px, tag_type);
		return CLAUSE_PARAGRAPH + voice_change_flag;
	case SSML_PARAGRAPH:
		if (ssml_sp->tag_type == SSML_SENTENCE) {
			// new paragraph implies end-of-sentence or end-of-paragraph
			voice_change_flag = GetVoiceAttributes(px, SSML_SENTENCE+SSML_CLOSE);
		}
		if (ssml_sp->tag_type == SSML_PARAGRAPH) {
			// new paragraph implies end-of-sentence or end-of-paragraph
			voice_change_flag |= GetVoiceAttributes(px, SSML_PARAGRAPH+SSML_CLOSE);
		}
		voice_change_flag |= GetVoiceAttributes(px, tag_type);
		return CLAUSE_PARAGRAPH + voice_change_flag;
	case SSML_SENTENCE + SSML_CLOSE:
		if (ssml_sp->tag_type == SSML_SENTENCE) {
			// end of a sentence which specified a language
			voice_change_flag = GetVoiceAttributes(px, tag_type);
		}
		return CLAUSE_PERIOD + voice_change_flag;
	case SSML_PARAGRAPH + SSML_CLOSE:
		if ((ssml_sp->tag_type == SSML_SENTENCE) || (ssml_sp->tag_type == SSML_PARAGRAPH)) {
			// End of a paragraph which specified a language.
			// (End-of-paragraph also implies end-of-sentence)
			return GetVoiceAttributes(px, tag_type) + CLAUSE_PARAGRAPH;
		}
		return CLAUSE_PARAGRAPH;
	}
	return 0;
}

static void RemoveChar(char *p)
{
	// Replace a UTF-8 character by spaces
	int c;

	memset(p, ' ', utf8_in(&c, p));
}

static MNEM_TAB xml_char_mnemonics[] = {
	{ "gt",   '>' },
	{ "lt",   0xe000 + '<' },   // private usage area, to avoid confusion with XML tag
	{ "amp",  '&' },
	{ "quot", '"' },
	{ "nbsp", ' ' },
	{ "apos", '\'' },
	{ NULL,   -1 }
};

int ReadClause(Translator *tr, char *buf, short *charix, int *charix_top, int n_buf, int *tone_type, char *voice_change)
{
	/* Find the end of the current clause.
	    Write the clause into  buf

	    returns: clause type (bits 0-7: pause x10mS, bits 8-11 intonation type)

	    Also checks for blank line (paragraph) as end-of-clause indicator.

	    Does not end clause for:
	        punctuation immediately followed by alphanumeric  eg.  1.23  !Speak  :path
	        repeated punctuation, eg.   ...   !!!
	 */

	int c1 = ' '; // current character
	int c2; // next character
	int cprev = ' '; // previous character
	int cprev2 = ' ';
	int c_next;
	int parag;
	int ix = 0;
	int j;
	int nl_count;
	int linelength = 0;
	int phoneme_mode = 0;
	int n_xml_buf;
	int terminator;
	int found;
	int any_alnum = 0;
	int self_closing;
	int punct_data = 0;
	int is_end_clause;
	int announced_punctuation = 0;
	int stressed_word = 0;
	int end_clause_after_tag = 0;
	int end_clause_index = 0;
	wchar_t xml_buf[N_XML_BUF+1];

	#define N_XML_BUF2 20
	char xml_buf2[N_XML_BUF2+2]; // for &<name> and &<number> sequences
	static char ungot_string[N_XML_BUF2+4];
	static int ungot_string_ix = -1;

	if (clear_skipping_text) {
		skipping_text = 0;
		clear_skipping_text = 0;
	}

	tr->phonemes_repeat_count = 0;
	tr->clause_upper_count = 0;
	tr->clause_lower_count = 0;
	*tone_type = 0;
	*voice_change = 0;

	if (ungot_word != NULL) {
		strcpy(buf, ungot_word);
		ix += strlen(ungot_word);
		ungot_word = NULL;
	}

	if (ungot_char2 != 0)
		c2 = ungot_char2;
	else
		c2 = GetC();

	while (!Eof() || (ungot_char != 0) || (ungot_char2 != 0) || (ungot_string_ix >= 0)) {
		if (!iswalnum(c1)) {
			if ((end_character_position > 0) && (count_characters > end_character_position)) {
				return CLAUSE_EOF;
			}

			if ((skip_characters > 0) && (count_characters >= skip_characters)) {
				// reached the specified start position
				// don't break a word
				clear_skipping_text = 1;
				skip_characters = 0;
				UngetC(c2);
				return CLAUSE_NONE;
			}
		}

		cprev2 = cprev;
		cprev = c1;
		c1 = c2;

		if (ungot_string_ix >= 0) {
			if (ungot_string[ungot_string_ix] == 0)
				ungot_string_ix = -1;
		}

		if ((ungot_string_ix == 0) && (ungot_char2 == 0))
			c1 = ungot_string[ungot_string_ix++];
		if (ungot_string_ix >= 0)
			c2 = ungot_string[ungot_string_ix++];
		else {
			c2 = GetC();

			if (Eof())
				c2 = ' ';
		}
		ungot_char2 = 0;

		if ((option_ssml) && (phoneme_mode == 0)) {
			if ((ssml_ignore_l_angle != '&') && (c1 == '&') && ((c2 == '#') || ((c2 >= 'a') && (c2 <= 'z')))) {
				n_xml_buf = 0;
				c1 = c2;
				while (!Eof() && (iswalnum(c1) || (c1 == '#')) && (n_xml_buf < N_XML_BUF2)) {
					xml_buf2[n_xml_buf++] = c1;
					c1 = GetC();
				}
				xml_buf2[n_xml_buf] = 0;
				c2 = GetC();
				sprintf(ungot_string, "%s%c%c", &xml_buf2[0], c1, c2);

				if (c1 == ';') {
					if (xml_buf2[0] == '#') {
						// character code number
						if (xml_buf2[1] == 'x')
							found = sscanf(&xml_buf2[2], "%x", (unsigned int *)(&c1));
						else
							found = sscanf(&xml_buf2[1], "%d", &c1);
					} else {
						if ((found = LookupMnem(xml_char_mnemonics, xml_buf2)) != -1) {
							c1 = found;
							if (c2 == 0)
								c2 = ' ';
						}
					}
				} else
					found = -1;

				if (found <= 0) {
					ungot_string_ix = 0;
					c1 = '&';
					c2 = ' ';
				}

				if ((c1 <= 0x20) && ((sayas_mode == SAYAS_SINGLE_CHARS) || (sayas_mode == SAYAS_KEY)))
					c1 += 0xe000; // move into unicode private usage area
			} else if ((c1 == '<') && (ssml_ignore_l_angle != '<')) {
				if ((c2 == '!') || (c2 == '?')) {
					// a comment, ignore until closing '<'  (or <?xml tag )
					while (!Eof() && (c1 != '>'))
						c1 = GetC();
					c2 = ' ';
				} else if ((c2 == '/') || iswalpha(c2)) {
					// check for space in the output buffer for embedded commands produced by the SSML tag
					if (ix > (n_buf - 20)) {
						// Perhaps not enough room, end the clause before the SSML tag
						UngetC(c2);
						ungot_char2 = c1;
						buf[ix] = ' ';
						buf[ix+1] = 0;
						return CLAUSE_NONE;
					}

					// SSML Tag
					n_xml_buf = 0;
					c1 = c2;
					while (!Eof() && (c1 != '>') && (n_xml_buf < N_XML_BUF)) {
						xml_buf[n_xml_buf++] = c1;
						c1 = GetC();
					}
					xml_buf[n_xml_buf] = 0;
					c2 = ' ';

					self_closing = 0;
					if (xml_buf[n_xml_buf-1] == '/') {
						// a self-closing tag
						xml_buf[n_xml_buf-1] = ' ';
						self_closing = 1;
					}

					terminator = ProcessSsmlTag(xml_buf, buf, &ix, n_buf, self_closing);

					if (terminator != 0) {
						if (end_clause_after_tag)
							ix = end_clause_index;

						buf[ix] = ' ';
						buf[ix++] = 0;

						if (terminator & CLAUSE_TYPE_VOICE_CHANGE)
							strcpy(voice_change, current_voice_id);
						return terminator;
					}
					c1 = ' ';
					c2 = GetC();
					continue;
				}
			}
		}
		ssml_ignore_l_angle = 0;

		if (ignore_text)
			continue;

		if ((c2 == '\n') && (option_linelength == -1)) {
			// single-line mode, return immediately on NL
			if ((terminator = clause_type_from_codepoint(c1)) == CLAUSE_NONE) {
				charix[ix] = count_characters - clause_start_char;
				*charix_top = ix;
				ix += utf8_out(c1, &buf[ix]);
				terminator = CLAUSE_PERIOD; // line doesn't end in punctuation, assume period
			}
			buf[ix] = ' ';
			buf[ix+1] = 0;
			return terminator;
		}

		if ((c1 == CTRL_EMBEDDED) || (c1 == ctrl_embedded)) {
			// an embedded command. If it's a voice change, end the clause
			if (c2 == 'V') {
				buf[ix++] = 0; // end the clause at this point
				while (!iswspace(c1 = GetC()) && !Eof() && (ix < (n_buf-1)))
					buf[ix++] = c1; // add voice name to end of buffer, after the text
				buf[ix++] = 0;
				return CLAUSE_VOICE;
			} else if (c2 == 'B') {
				// set the punctuation option from an embedded command
				//  B0     B1     B<punct list><space>
				strcpy(&buf[ix], "   ");
				ix += 3;

				if ((c2 = GetC()) == '0')
					option_punctuation = 0;
				else {
					option_punctuation = 1;
					option_punctlist[0] = 0;
					if (c2 != '1') {
						// a list of punctuation characters to be spoken, terminated by space
						j = 0;
						while (!iswspace(c2) && !Eof()) {
							option_punctlist[j++] = c2;
							c2 = GetC();
							buf[ix++] = ' ';
						}
						option_punctlist[j] = 0; // terminate punctuation list
						option_punctuation = 2;
					}
				}
				c2 = GetC();
				continue;
			}
		}

		linelength++;

		if ((j = lookupwchar2(tr->chars_ignore, c1)) != 0) {
			if (j == 1) {
				// ignore this character (eg. zero-width-non-joiner U+200C)
				continue;
			}
			c1 = j; // replace the character
		}

		if (iswalnum(c1))
			any_alnum = 1;
		else {
			if (stressed_word) {
				stressed_word = 0;
				c1 = CHAR_EMPHASIS; // indicate this word is stressed
				UngetC(c2);
				c2 = ' ';
			}

			if (c1 == 0xf0b)
				c1 = ' '; // Tibet inter-syllabic mark, ?? replace by space ??

			if (iswspace(c1)) {
				char *p_word;

				if (tr->translator_name == 0x6a626f) {
					// language jbo : lojban
					// treat "i" or ".i" as end-of-sentence
					p_word = &buf[ix-1];
					if (p_word[0] == 'i') {
						if (p_word[-1] == '.')
							p_word--;
						if (p_word[-1] == ' ') {
							ungot_word = "i ";
							UngetC(c2);
							p_word[0] = 0;
							return CLAUSE_PERIOD;
						}
					}
				}
			}

			if (c1 == 0xd4d) {
				// Malayalam virama, check if next character is Zero-width-joiner
				if (c2 == 0x200d)
					c1 = 0xd4e; // use this unofficial code for chillu-virama
			}
		}

		if (iswupper(c1)) {
			tr->clause_upper_count++;
			if ((option_capitals == 2) && (sayas_mode == 0) && !iswupper(cprev)) {
				char text_buf[40];
				char text_buf2[30];
				if (LookupSpecial(tr, "_cap", text_buf2) != NULL) {
					sprintf(text_buf, "%s", text_buf2);
					j = strlen(text_buf);
					if ((ix + j) < n_buf) {
						strcpy(&buf[ix], text_buf);
						ix += j;
					}
				}
			}
		} else if (iswalpha(c1))
			tr->clause_lower_count++;

		if (option_phoneme_input) {
			if (phoneme_mode > 0)
				phoneme_mode--;
			else if ((c1 == '[') && (c2 == '['))
				phoneme_mode = -1; // input is phoneme mnemonics, so don't look for punctuation
			else if ((c1 == ']') && (c2 == ']'))
				phoneme_mode = 2; // set phoneme_mode to zero after the next two characters
		}

		if (c1 == '\n') {
			parag = 0;

			// count consecutive newlines, ignoring other spaces
			while (!Eof() && iswspace(c2)) {
				if (c2 == '\n')
					parag++;
				c2 = GetC();
			}
			if (parag > 0) {
				// 2nd newline, assume paragraph
				UngetC(c2);

				if (end_clause_after_tag)
					RemoveChar(&buf[end_clause_index]); // delete clause-end punctiation
				buf[ix] = ' ';
				buf[ix+1] = 0;
				if (parag > 3)
					parag = 3;
				if (option_ssml) parag = 1;
				return (CLAUSE_PARAGRAPH-30) + 30*parag; // several blank lines, longer pause
			}

			if (linelength <= option_linelength) {
				// treat lines shorter than a specified length as end-of-clause
				UngetC(c2);
				buf[ix] = ' ';
				buf[ix+1] = 0;
				return CLAUSE_COLON;
			}

			linelength = 0;
		}

		announced_punctuation = 0;

		if ((phoneme_mode == 0) && (sayas_mode == 0)) {
			is_end_clause = 0;

			if (end_clause_after_tag) {
				// Because of an xml tag, we are waiting for the
				// next non-blank character to decide whether to end the clause
				// i.e. is dot followed by an upper-case letter?

				if (!iswspace(c1)) {
					if (!IsAlpha(c1) || !iswlower(c1)) {
						UngetC(c2);
						ungot_char2 = c1;
						buf[end_clause_index] = ' '; // delete the end-clause punctuation
						buf[end_clause_index+1] = 0;
						return end_clause_after_tag;
					}
					end_clause_after_tag = 0;
				}
			}

			if ((c1 == '.') && (c2 == '.')) {
				while ((c_next = GetC()) == '.') {
					// 3 or more dots, replace by elipsis
					c1 = 0x2026;
					c2 = ' ';
				}
				if (c1 == 0x2026)
					c2 = c_next;
				else
					UngetC(c_next);
			}

			punct_data = 0;
			if ((punct_data = clause_type_from_codepoint(c1)) != CLAUSE_NONE) {
				if (punct_data & CLAUSE_PUNCTUATION_IN_WORD) {
					// Armenian punctuation inside a word
					stressed_word = 1;
					*tone_type = punct_data >> 12 & 0xf; // override the end-of-sentence type
					continue;
				}

				if ((iswspace(c2) || (punct_data & CLAUSE_OPTIONAL_SPACE_AFTER) || IsBracket(c2) || (c2 == '?') || Eof() || (c2 == ctrl_embedded))) { // don't check for '-' because it prevents recognizing ':-)'
					// note: (c2='?') is for when a smart-quote has been replaced by '?'
					is_end_clause = 1;
				}
			}

			// don't announce punctuation for the alternative text inside inside <audio> ... </audio>
			if (c1 == 0xe000+'<')  c1 = '<';
			if (option_punctuation && iswpunct(c1) && (audio_text == 0)) {
				// option is set to explicitly speak punctuation characters
				// if a list of allowed punctuation has been set up, check whether the character is in it
				if ((option_punctuation == 1) || (wcschr(option_punctlist, c1) != NULL)) {
					tr->phonemes_repeat_count = 0;
					if ((terminator = AnnouncePunctuation(tr, c1, &c2, buf, &ix, is_end_clause)) >= 0)
						return terminator;
					announced_punctuation = c1;
				}
			}

			if ((punct_data & CLAUSE_SPEAK_PUNCTUATION_NAME) && (announced_punctuation == 0)) {
				// used for elipsis (and 3 dots) if a pronunciation for elipsis is given in *_list
				char *p2;

				p2 = &buf[ix];
				sprintf(p2, "%s", LookupCharName(tr, c1, 1));
				if (p2[0] != 0) {
					ix += strlen(p2);
					announced_punctuation = c1;
					punct_data = punct_data & ~CLAUSE_INTONATION_TYPE; // change intonation type to 0 (full-stop)
				}
			}

			if (is_end_clause) {
				nl_count = 0;
				c_next = c2;

				if (iswspace(c_next)) {
					while (!Eof() && iswspace(c_next)) {
						if (c_next == '\n')
							nl_count++;
						c_next = GetC(); // skip past space(s)
					}
				}

				if ((c1 == '.') && (nl_count < 2))
					punct_data |= CLAUSE_DOT_AFTER_LAST_WORD;

				if (nl_count == 0) {
					if ((c1 == ',') && (cprev == '.') && (tr->translator_name == L('h', 'u')) && iswdigit(cprev2) && (iswdigit(c_next) || (iswlower(c_next)))) {
						// lang=hu, fix for ordinal numbers, eg:  "december 2., szerda", ignore ',' after ordinal number
						c1 = CHAR_COMMA_BREAK;
						is_end_clause = 0;
					}

					if (c1 == '.') {
						if ((tr->langopts.numbers & NUM_ORDINAL_DOT) &&
						    (iswdigit(cprev) || (IsRomanU(cprev) && (IsRomanU(cprev2) || iswspace(cprev2))))) { // lang=hu
							// dot after a number indicates an ordinal number
							if (!iswdigit(cprev))
								is_end_clause = 0; // Roman number followed by dot
							else if (iswlower(c_next) || (c_next == '-')) // hyphen is needed for lang-hu (eg. 2.-kal)
								is_end_clause = 0; // only if followed by lower-case, (or if there is a XML tag)
						} else if (c_next == '\'')
							is_end_clause = 0;    // eg. u.s.a.'s
						if (iswlower(c_next)) {
							// next word has no capital letter, this dot is probably from an abbreviation
							is_end_clause = 0;
						}
						if (any_alnum == 0) {
							// no letters or digits yet, so probably not a sentence terminator
							// Here, dot is followed by space or bracket
							c1 = ' ';
							is_end_clause = 0;
						}
					} else {
						if (any_alnum == 0) {
							// no letters or digits yet, so probably not a sentence terminator
							is_end_clause = 0;
						}
					}

					if (is_end_clause && (c1 == '.') && (c_next == '<') && option_ssml) {
						// wait until after the end of the xml tag, then look for upper-case letter
						is_end_clause = 0;
						end_clause_index = ix;
						end_clause_after_tag = punct_data;
					}
				}

				if (is_end_clause) {
					UngetC(c_next);
					buf[ix] = ' ';
					buf[ix+1] = 0;

					if (iswdigit(cprev) && !IsAlpha(c_next)) // ????
						punct_data &= ~CLAUSE_DOT_AFTER_LAST_WORD;
					if (nl_count > 1) {
						if ((punct_data == CLAUSE_QUESTION) || (punct_data == CLAUSE_EXCLAMATION))
							return punct_data + 35; // with a longer pause
						return CLAUSE_PARAGRAPH;
					}
					return punct_data; // only recognise punctuation if followed by a blank or bracket/quote
				} else if (!Eof()) {
					if (iswspace(c2))
						UngetC(c_next);
				}
			}
		}

		if (speech_parameters[espeakSILENCE] == 1)
			continue;

		if (c1 == announced_punctuation) {
			// This character has already been announced, so delete it so that it isn't spoken a second time.
			// Unless it's a hyphen or apostrophe (which is used by TranslateClause() )
			if (IsBracket(c1))
				c1 = 0xe000 + '('; // Unicode private useage area.  So TranslateRules() knows the bracket name has been spoken
			else if (c1 != '-')
				c1 = ' ';
		}

		j = ix+1;

		if (c1 == 0xe000 + '<') c1 = '<';

		ix += utf8_out(c1, &buf[ix]);
		if (!iswspace(c1) && !IsBracket(c1)) {
			charix[ix] = count_characters - clause_start_char;
			while (j < ix)
				charix[j++] = -1; // subsequent bytes of a multibyte character
		}
		*charix_top = ix;

		if (((ix > (n_buf-75)) && !IsAlpha(c1) && !iswdigit(c1))  ||  (ix >= (n_buf-4))) {
			// clause too long, getting near end of buffer, so break here
			// try to break at a word boundary (unless we actually reach the end of buffer).
			// (n_buf-4) is to allow for 3 bytes of multibyte character plus terminator.
			buf[ix] = ' ';
			buf[ix+1] = 0;
			UngetC(c2);
			return CLAUSE_NONE;
		}
	}

	if (stressed_word)
		ix += utf8_out(CHAR_EMPHASIS, &buf[ix]);
	if (end_clause_after_tag)
		RemoveChar(&buf[end_clause_index]); // delete clause-end punctiation
	buf[ix] = ' ';
	buf[ix+1] = 0;
	return CLAUSE_EOF; // end of file
}

void InitNamedata(void)
{
	namedata_ix = 0;
	if (namedata != NULL) {
		free(namedata);
		namedata = NULL;
		n_namedata = 0;
	}
}

void InitText2(void)
{
	int param;

	ungot_char = 0;
	ungot_char2 = 0;

	n_ssml_stack = 1;
	n_param_stack = 1;
	ssml_stack[0].tag_type = 0;

	for (param = 0; param < N_SPEECH_PARAM; param++)
		speech_parameters[param] = param_stack[0].parameter[param]; // set all speech parameters to defaults

	option_punctuation = speech_parameters[espeakPUNCTUATION];
	option_capitals = speech_parameters[espeakCAPITALS];

	current_voice_id[0] = 0;

	ignore_text = 0;
	audio_text = 0;
	clear_skipping_text = 0;
	count_characters = -1;
	sayas_mode = 0;

	xmlbase = NULL;
}