| 1 | """edge-tts backend for narration audio generation.""" |
| 2 | |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | import os |
| 6 | import re |
| 7 | from dataclasses import dataclass |
| 8 | from pathlib import Path |
| 9 | |
| 10 | from tts_backends.backend_common import publish_staged_pair, temporary_path |
| 11 | |
| 12 | |
| 13 | DEFAULT_SUBTITLE_MAX_CHARS = 20 |
| 14 | DEFAULT_BOUNDARY_OVERLAP_TOLERANCE_MS = 100 |
| 15 | _TICKS_PER_MILLISECOND = 10_000 |
| 16 | _SENTENCE_END = frozenset("。！？!?") |
| 17 | _CLAUSE_END = frozenset("，,；;：:") |
| 18 | _CLOSING_PUNCTUATION = frozenset('”’」』）》)"\'') |
| 19 | |
| 20 | |
| 21 | @dataclass(frozen=True) |
| 22 | class _MappedWord: |
| 23 | start: int |
| 24 | end: int |
| 25 | source_start: int |
| 26 | source_end: int |
| 27 | |
| 28 | |
| 29 | @dataclass(frozen=True) |
| 30 | class _SubtitleCue: |
| 31 | start: int |
| 32 | end: int |
| 33 | text: str |
| 34 | |
| 35 | |
| 36 | COMMON_VOICES = [ |
| 37 | ("zh-CN", "zh-CN-XiaoxiaoNeural", "女声，普通话，清晰自然，默认推荐"), |
| 38 | ("zh-CN", "zh-CN-XiaoyiNeural", "女声，普通话，明亮"), |
| 39 | ("zh-CN", "zh-CN-YunjianNeural", "男声，普通话，稳重"), |
| 40 | ("zh-CN", "zh-CN-YunxiNeural", "男声，普通话，年轻"), |
| 41 | ("zh-CN", "zh-CN-YunxiaNeural", "男声，普通话，少年感"), |
| 42 | ("zh-CN", "zh-CN-YunyangNeural", "男声，普通话，播报感"), |
| 43 | ("zh-HK", "zh-HK-HiuGaaiNeural", "女声，粤语"), |
| 44 | ("zh-HK", "zh-HK-WanLungNeural", "男声，粤语"), |
| 45 | ("zh-TW", "zh-TW-HsiaoChenNeural", "女声，台湾普通话"), |
| 46 | ("zh-TW", "zh-TW-YunJheNeural", "男声，台湾普通话"), |
| 47 | ("en-US", "en-US-JennyNeural", "女声，美式英语"), |
| 48 | ("en-US", "en-US-GuyNeural", "男声，美式英语"), |
| 49 | ("en-GB", "en-GB-SoniaNeural", "女声，英式英语"), |
| 50 | ("en-GB", "en-GB-RyanNeural", "男声，英式英语"), |
| 51 | ] |
| 52 | |
| 53 | |
| 54 | def edge_output_extension() -> str: |
| 55 | return ".mp3" |
| 56 | |
| 57 | |
| 58 | def normalize_rate(rate: str) -> str: |
| 59 | """Normalize a user-provided rate into edge-tts format.""" |
| 60 | value = rate.strip() |
| 61 | if not value: |
| 62 | return "+0%" |
| 63 | if value.endswith("%"): |
| 64 | if value[0] not in "+-": |
| 65 | return f"+{value}" |
| 66 | return value |
| 67 | if re.fullmatch(r"[+-]?\d+", value): |
| 68 | number = int(value) |
| 69 | return f"{number:+d}%" |
| 70 | return value |
| 71 | |
| 72 | |
| 73 | async def generate( |
| 74 | text: str, |
| 75 | output_path: Path, |
| 76 | *, |
| 77 | voice: str, |
| 78 | rate: str, |
| 79 | subtitle_path: Path | None = None, |
| 80 | subtitle_max_chars: int = DEFAULT_SUBTITLE_MAX_CHARS, |
| 81 | ) -> None: |
| 82 | """Generate narration audio and, when requested, its compact SRT.""" |
| 83 | if subtitle_path is not None: |
| 84 | await _generate_with_subtitles( |
| 85 | text, |
| 86 | output_path, |
| 87 | subtitle_path, |
| 88 | voice=voice, |
| 89 | rate=rate, |
| 90 | max_chars=subtitle_max_chars, |
| 91 | ) |
| 92 | return |
| 93 | |
| 94 | try: |
| 95 | import edge_tts |
| 96 | except ImportError as exc: |
| 97 | raise RuntimeError( |
| 98 | "Missing dependency `edge-tts`. Install it with: " |
| 99 | "python3 -m pip install edge-tts" |
| 100 | ) from exc |
| 101 | |
| 102 | communicate = edge_tts.Communicate(text, voice=voice, rate=normalize_rate(rate)) |
| 103 | await communicate.save(str(output_path)) |
| 104 | |
| 105 | |
| 106 | def _text_key(text: str) -> str: |
| 107 | return "".join(character.casefold() for character in text if character.isalnum()) |
| 108 | |
| 109 | |
| 110 | def _source_key_positions(text: str) -> tuple[str, list[int]]: |
| 111 | key: list[str] = [] |
| 112 | positions: list[int] = [] |
| 113 | for index, character in enumerate(text): |
| 114 | if not character.isalnum(): |
| 115 | continue |
| 116 | normalized = character.casefold() |
| 117 | key.extend(normalized) |
| 118 | positions.extend([index] * len(normalized)) |
| 119 | return "".join(key), positions |
| 120 | |
| 121 | |
| 122 | def _map_word_boundaries( |
| 123 | text: str, |
| 124 | boundaries: list[dict], |
| 125 | provider_label: str, |
| 126 | ) -> list[_MappedWord]: |
| 127 | source_key, source_positions = _source_key_positions(text) |
| 128 | boundary_keys = [_text_key(boundary["text"]) for boundary in boundaries] |
| 129 | boundary_key = "".join(boundary_keys) |
| 130 | if not source_key or source_key != boundary_key: |
| 131 | raise RuntimeError( |
| 132 | f"{provider_label} word boundaries could not be aligned with the narration text; " |
| 133 | "subtitle timing was not generated" |
| 134 | ) |
| 135 | |
| 136 | mapped: list[_MappedWord] = [] |
| 137 | key_offset = 0 |
| 138 | for boundary, word_key in zip(boundaries, boundary_keys): |
| 139 | if not word_key: |
| 140 | continue |
| 141 | key_end = key_offset + len(word_key) |
| 142 | mapped.append( |
| 143 | _MappedWord( |
| 144 | start=boundary["offset"], |
| 145 | end=boundary["offset"] + boundary["duration"], |
| 146 | source_start=source_positions[key_offset], |
| 147 | source_end=source_positions[key_end - 1] + 1, |
| 148 | ) |
| 149 | ) |
| 150 | key_offset = key_end |
| 151 | return mapped |
| 152 | |
| 153 | |
| 154 | def _trim_span(text: str, start: int, end: int) -> tuple[int, int]: |
| 155 | while start < end and text[start].isspace(): |
| 156 | start += 1 |
| 157 | while end > start and text[end - 1].isspace(): |
| 158 | end -= 1 |
| 159 | return start, end |
| 160 | |
| 161 | |
| 162 | def _display_length(text: str, start: int, end: int) -> int: |
| 163 | return sum(not character.isspace() for character in text[start:end]) |
| 164 | |
| 165 | |
| 166 | def _is_sentence_end(text: str, index: int) -> bool: |
| 167 | character = text[index] |
| 168 | if character in _SENTENCE_END: |
| 169 | return True |
| 170 | if character != ".": |
| 171 | return False |
| 172 | previous = text[index - 1] if index else "" |
| 173 | following = text[index + 1] if index + 1 < len(text) else "" |
| 174 | return not (previous.isdigit() and following.isdigit()) |
| 175 | |
| 176 | |
| 177 | def _sentence_spans(text: str) -> list[tuple[int, int]]: |
| 178 | spans: list[tuple[int, int]] = [] |
| 179 | start = 0 |
| 180 | index = 0 |
| 181 | while index < len(text): |
| 182 | if not _is_sentence_end(text, index): |
| 183 | index += 1 |
| 184 | continue |
| 185 | end = index + 1 |
| 186 | while end < len(text) and text[end] in _CLOSING_PUNCTUATION: |
| 187 | end += 1 |
| 188 | span = _trim_span(text, start, end) |
| 189 | if span[0] < span[1]: |
| 190 | spans.append(span) |
| 191 | start = end |
| 192 | index = end |
| 193 | span = _trim_span(text, start, len(text)) |
| 194 | if span[0] < span[1]: |
| 195 | spans.append(span) |
| 196 | return spans |
| 197 | |
| 198 | |
| 199 | def _hard_split_span( |
| 200 | text: str, |
| 201 | span: tuple[int, int], |
| 202 | words: list[_MappedWord], |
| 203 | max_chars: int, |
| 204 | ) -> list[tuple[int, int]]: |
| 205 | start, end = span |
| 206 | parts: list[tuple[int, int]] = [] |
| 207 | while _display_length(text, start, end) > max_chars: |
| 208 | remaining_length = _display_length(text, start, end) |
| 209 | remaining_parts = (remaining_length + max_chars - 1) // max_chars |
| 210 | target_length = (remaining_length + remaining_parts - 1) // remaining_parts |
| 211 | candidates = [ |
| 212 | (word.source_end, _display_length(text, start, word.source_end)) |
| 213 | for word in words |
| 214 | if start < word.source_end < end |
| 215 | and _display_length(text, start, word.source_end) <= max_chars |
| 216 | ] |
| 217 | if candidates: |
| 218 | split_at, _ = min( |
| 219 | candidates, |
| 220 | key=lambda candidate: ( |
| 221 | abs(candidate[1] - target_length), |
| 222 | -candidate[1], |
| 223 | ), |
| 224 | ) |
| 225 | else: |
| 226 | split_at = next( |
| 227 | ( |
| 228 | word.source_end |
| 229 | for word in words |
| 230 | if start < word.source_end < end |
| 231 | ), |
| 232 | end, |
| 233 | ) |
| 234 | if split_at >= end: |
| 235 | break |
| 236 | part = _trim_span(text, start, split_at) |
| 237 | if part[0] < part[1]: |
| 238 | parts.append(part) |
| 239 | start = split_at |
| 240 | part = _trim_span(text, start, end) |
| 241 | if part[0] < part[1]: |
| 242 | parts.append(part) |
| 243 | return parts |
| 244 | |
| 245 | |
| 246 | def _split_sentence_span( |
| 247 | text: str, |
| 248 | sentence: tuple[int, int], |
| 249 | words: list[_MappedWord], |
| 250 | max_chars: int, |
| 251 | ) -> list[tuple[int, int]]: |
| 252 | if _display_length(text, *sentence) <= max_chars: |
| 253 | return [sentence] |
| 254 | |
| 255 | start, end = sentence |
| 256 | clauses: list[tuple[int, int]] = [] |
| 257 | clause_start = start |
| 258 | for index in range(start, end): |
| 259 | if text[index] not in _CLAUSE_END: |
| 260 | continue |
| 261 | clause = _trim_span(text, clause_start, index + 1) |
| 262 | if clause[0] < clause[1]: |
| 263 | clauses.append(clause) |
| 264 | clause_start = index + 1 |
| 265 | clause = _trim_span(text, clause_start, end) |
| 266 | if clause[0] < clause[1]: |
| 267 | clauses.append(clause) |
| 268 | |
| 269 | atoms = [ |
| 270 | part |
| 271 | for clause in clauses |
| 272 | for part in _hard_split_span(text, clause, words, max_chars) |
| 273 | ] |
| 274 | merged: list[tuple[int, int]] = [] |
| 275 | for atom in atoms: |
| 276 | if not merged: |
| 277 | merged.append(atom) |
| 278 | continue |
| 279 | candidate = (merged[-1][0], atom[1]) |
| 280 | if _display_length(text, *candidate) <= max_chars: |
| 281 | merged[-1] = candidate |
| 282 | else: |
| 283 | merged.append(atom) |
| 284 | return merged |
| 285 | |
| 286 | |
| 287 | def _clamp_small_overlaps( |
| 288 | cues: list[_SubtitleCue], |
| 289 | *, |
| 290 | tolerance_ms: int = DEFAULT_BOUNDARY_OVERLAP_TOLERANCE_MS, |
| 291 | provider_label: str, |
| 292 | ) -> list[_SubtitleCue]: |
| 293 | tolerance = tolerance_ms * _TICKS_PER_MILLISECOND |
| 294 | normalized: list[_SubtitleCue] = [] |
| 295 | for cue in cues: |
| 296 | if normalized and cue.start < normalized[-1].end: |
| 297 | overlap = normalized[-1].end - cue.start |
| 298 | if overlap > tolerance: |
| 299 | overlap_ms = overlap / _TICKS_PER_MILLISECOND |
| 300 | raise RuntimeError( |
| 301 | f"{provider_label} returned {overlap_ms:g} ms of overlapping " |
| 302 | "word-boundary timing; audio and subtitles were not published" |
| 303 | ) |
| 304 | cue = _SubtitleCue( |
| 305 | start=normalized[-1].end, |
| 306 | end=cue.end, |
| 307 | text=cue.text, |
| 308 | ) |
| 309 | if cue.end <= cue.start: |
| 310 | raise RuntimeError( |
| 311 | f"{provider_label} returned an invalid subtitle timing interval; " |
| 312 | "audio and subtitles were not published" |
| 313 | ) |
| 314 | normalized.append(cue) |
| 315 | return normalized |
| 316 | |
| 317 | |
| 318 | def _subtitle_cues( |
| 319 | text: str, |
| 320 | boundaries: list[dict], |
| 321 | max_chars: int, |
| 322 | provider_label: str, |
| 323 | ) -> list[_SubtitleCue]: |
| 324 | if max_chars < 1: |
| 325 | raise ValueError("subtitle_max_chars must be at least 1") |
| 326 | words = _map_word_boundaries(text, boundaries, provider_label) |
| 327 | spans = [ |
| 328 | span |
| 329 | for sentence in _sentence_spans(text) |
| 330 | for span in _split_sentence_span(text, sentence, words, max_chars) |
| 331 | ] |
| 332 | |
| 333 | pending: list[tuple[int, int, str]] = [] |
| 334 | assigned_word_indexes: list[int] = [] |
| 335 | for start, end in spans: |
| 336 | matching = [ |
| 337 | (index, word) |
| 338 | for index, word in enumerate(words) |
| 339 | if word.source_start >= start and word.source_end <= end |
| 340 | ] |
| 341 | if not matching: |
| 342 | continue |
| 343 | cue_text = re.sub(r"\s+", " ", text[start:end]).strip() |
| 344 | assigned_word_indexes.extend(index for index, _ in matching) |
| 345 | pending.append((matching[0][1].start, matching[-1][1].end, cue_text)) |
| 346 | |
| 347 | if assigned_word_indexes != list(range(len(words))): |
| 348 | raise RuntimeError( |
| 349 | f"{provider_label} word boundaries crossed subtitle split points; " |
| 350 | "subtitle timing was not generated" |
| 351 | ) |
| 352 | if not pending: |
| 353 | raise RuntimeError(f"{provider_label} produced no timed subtitle cues") |
| 354 | |
| 355 | cues: list[_SubtitleCue] = [] |
| 356 | for index, (start, word_end, cue_text) in enumerate(pending): |
| 357 | next_start = pending[index + 1][0] if index + 1 < len(pending) else None |
| 358 | end = next_start if next_start is not None and next_start > word_end else word_end |
| 359 | cues.append(_SubtitleCue(start=start, end=end, text=cue_text)) |
| 360 | cues = _clamp_small_overlaps(cues, provider_label=provider_label) |
| 361 | |
| 362 | source_text = re.sub(r"\s+", "", text) |
| 363 | subtitle_text = re.sub(r"\s+", "", "".join(cue.text for cue in cues)) |
| 364 | if subtitle_text != source_text: |
| 365 | raise RuntimeError( |
| 366 | f"{provider_label} subtitle text does not match the narration text; " |
| 367 | "audio and subtitles were not published" |
| 368 | ) |
| 369 | if any(_display_length(cue.text, 0, len(cue.text)) > max_chars for cue in cues): |
| 370 | raise RuntimeError( |
| 371 | f"A single {provider_label} word boundary exceeds the subtitle character limit; " |
| 372 | "audio and subtitles were not published" |
| 373 | ) |
| 374 | return cues |
| 375 | |
| 376 | |
| 377 | def _srt_timestamp(ticks: int) -> str: |
| 378 | total_milliseconds = round(ticks / 10_000) |
| 379 | hours, remainder = divmod(total_milliseconds, 3_600_000) |
| 380 | minutes, remainder = divmod(remainder, 60_000) |
| 381 | seconds, milliseconds = divmod(remainder, 1_000) |
| 382 | return f"{hours:02d}:{minutes:02d}:{seconds:02d},{milliseconds:03d}" |
| 383 | |
| 384 | |
| 385 | def _format_srt(cues: list[_SubtitleCue]) -> str: |
| 386 | blocks = [ |
| 387 | ( |
| 388 | f"{index}\n" |
| 389 | f"{_srt_timestamp(cue.start)} --> {_srt_timestamp(cue.end)}\n" |
| 390 | f"{cue.text}" |
| 391 | ) |
| 392 | for index, cue in enumerate(cues, 1) |
| 393 | ] |
| 394 | return "\n\n".join(blocks) + "\n" |
| 395 | |
| 396 | |
| 397 | def format_word_timed_srt( |
| 398 | text: str, |
| 399 | boundaries: list[dict], |
| 400 | max_chars: int = DEFAULT_SUBTITLE_MAX_CHARS, |
| 401 | *, |
| 402 | provider_label: str = "Edge TTS", |
| 403 | ) -> str: |
| 404 | """Regroup provider word timings into compact, text-faithful SRT cues.""" |
| 405 | return _format_srt( |
| 406 | _subtitle_cues( |
| 407 | text, |
| 408 | boundaries, |
| 409 | max_chars, |
| 410 | provider_label, |
| 411 | ) |
| 412 | ) |
| 413 | |
| 414 | |
| 415 | async def _generate_with_subtitles( |
| 416 | text: str, |
| 417 | output_path: Path, |
| 418 | subtitle_path: Path, |
| 419 | *, |
| 420 | voice: str, |
| 421 | rate: str, |
| 422 | max_chars: int, |
| 423 | ) -> None: |
| 424 | """Generate one MP3 and compact SRT from the same Edge word-timing stream.""" |
| 425 | try: |
| 426 | import edge_tts |
| 427 | except ImportError as exc: |
| 428 | raise RuntimeError( |
| 429 | "Missing dependency `edge-tts`. Install it with: " |
| 430 | "python3 -m pip install edge-tts" |
| 431 | ) from exc |
| 432 | |
| 433 | communicate = edge_tts.Communicate( |
| 434 | text, |
| 435 | voice=voice, |
| 436 | rate=normalize_rate(rate), |
| 437 | boundary="WordBoundary", |
| 438 | ) |
| 439 | audio_descriptor = -1 |
| 440 | subtitle_descriptor = -1 |
| 441 | staged_audio: Path | None = None |
| 442 | staged_subtitle: Path | None = None |
| 443 | boundaries: list[dict] = [] |
| 444 | received_audio = False |
| 445 | try: |
| 446 | audio_descriptor, staged_audio = temporary_path(output_path, ".tmp") |
| 447 | subtitle_descriptor, staged_subtitle = temporary_path(subtitle_path, ".tmp") |
| 448 | |
| 449 | audio_stream = os.fdopen(audio_descriptor, "wb") |
| 450 | audio_descriptor = -1 |
| 451 | with audio_stream: |
| 452 | async for chunk in communicate.stream(): |
| 453 | if chunk["type"] == "audio": |
| 454 | audio_stream.write(chunk["data"]) |
| 455 | received_audio = True |
| 456 | elif chunk["type"] == "WordBoundary": |
| 457 | boundaries.append(chunk) |
| 458 | audio_stream.flush() |
| 459 | os.fsync(audio_stream.fileno()) |
| 460 | |
| 461 | if not received_audio: |
| 462 | raise RuntimeError("Edge TTS returned no audio data") |
| 463 | if not boundaries: |
| 464 | raise RuntimeError("Edge TTS returned no word-boundary timing") |
| 465 | subtitle_text = format_word_timed_srt(text, boundaries, max_chars) |
| 466 | |
| 467 | subtitle_stream = os.fdopen( |
| 468 | subtitle_descriptor, |
| 469 | "w", |
| 470 | encoding="utf-8", |
| 471 | newline="\n", |
| 472 | ) |
| 473 | subtitle_descriptor = -1 |
| 474 | with subtitle_stream: |
| 475 | subtitle_stream.write(subtitle_text) |
| 476 | subtitle_stream.flush() |
| 477 | os.fsync(subtitle_stream.fileno()) |
| 478 | assert staged_audio is not None |
| 479 | assert staged_subtitle is not None |
| 480 | publish_staged_pair( |
| 481 | staged_audio, |
| 482 | output_path, |
| 483 | staged_subtitle, |
| 484 | subtitle_path, |
| 485 | ) |
| 486 | finally: |
| 487 | if audio_descriptor >= 0: |
| 488 | os.close(audio_descriptor) |
| 489 | if subtitle_descriptor >= 0: |
| 490 | os.close(subtitle_descriptor) |
| 491 | if staged_audio is not None: |
| 492 | staged_audio.unlink(missing_ok=True) |
| 493 | if staged_subtitle is not None: |
| 494 | staged_subtitle.unlink(missing_ok=True) |
| 495 | |
| 496 | |
| 497 | def print_common_voices() -> None: |
| 498 | print("Common edge-tts voices:") |
| 499 | print("Locale Voice Notes") |
| 500 | print("------ ---------------------------- ----------------") |
| 501 | for locale, voice, notes in COMMON_VOICES: |
| 502 | print(f"{locale:<8} {voice:<29} {notes}") |
| 503 | |
| 504 | |
| 505 | async def print_voices(locale: str | None = None) -> None: |
| 506 | try: |
| 507 | import edge_tts |
| 508 | except ImportError as exc: |
| 509 | raise RuntimeError( |
| 510 | "Missing dependency `edge-tts`. Install it with: " |
| 511 | "python3 -m pip install edge-tts" |
| 512 | ) from exc |
| 513 | |
| 514 | manager = await edge_tts.VoicesManager.create() |
| 515 | voices = manager.voices |
| 516 | if locale: |
| 517 | voices = [voice for voice in voices if voice.get("Locale") == locale] |
| 518 | for voice in sorted(voices, key=lambda item: (item.get("Locale", ""), item.get("ShortName", ""))): |
| 519 | short_name = voice.get("ShortName", "") |
| 520 | voice_locale = voice.get("Locale", "") |
| 521 | gender = voice.get("Gender", "") |
| 522 | friendly = voice.get("FriendlyName", "") |
| 523 | print(f"{voice_locale:<8} {short_name:<34} {gender:<8} {friendly}") |
| 524 |