Spaces:
Running
Running
| """Pure helpers of the LFM3 Space (no model, no network): text pieces and voices. | |
| Kept apart from app.py so they can be unit-tested without torch or transformers. | |
| """ | |
| import re | |
| FIRST_PIECE_MAX = 60 # first sentence longer than this is cut at its first clause | |
| _SENTENCE = re.compile(r"(?<=[.!?…])\s+") | |
| _CLAUSE = re.compile(r"(?<=[,;:])\s+") | |
| # The Kokoro voice ids the app already sends (af_heart, bm_george, ...) are | |
| # accepted, so the same client can be pointed at either Space. LFM3 has no | |
| # named voices: only a language/locale and a female or male voice. | |
| _KOKORO_LANG = {"a": "American English", "b": "British English"} | |
| _GENDER = {"f": "female", "m": "male"} | |
| _KOKORO_VOICE = re.compile(r"^([ab])([fm])_[a-z]+$") | |
| def split_for_streaming(text: str) -> list: | |
| """Sentences, with the first one cut at its first clause when long, so | |
| the first audio arrives after synthesizing only a few words.""" | |
| sentences = [s.strip() for s in _SENTENCE.split(text) if s.strip()] | |
| if not sentences: | |
| return [] | |
| first = sentences[0] | |
| if len(first) > FIRST_PIECE_MAX: | |
| parts = _CLAUSE.split(first, maxsplit=1) | |
| if len(parts) == 2 and len(parts[0]) >= 12 and len(parts[1]) >= 12: | |
| return [parts[0], parts[1]] + sentences[1:] | |
| return sentences | |
| def parse_voice(voice: str): | |
| """(language, gender) for a voice id, or None when it is not supported. | |
| "female" / "male" mean American English; "af_heart"-style ids map the | |
| first letter to American (a) or British (b) English and the second to the | |
| gender. Portuguese Kokoro voices (pf_, pm_) are not mapped on purpose: | |
| this Space is for comparing English.""" | |
| voice = (voice or "").strip().lower() | |
| if voice in ("female", "male"): | |
| return "American English", voice | |
| match = _KOKORO_VOICE.match(voice) | |
| if match: | |
| return _KOKORO_LANG[match.group(1)], _GENDER[match.group(2)] | |
| return None | |
| def system_prompt(language: str, gender: str) -> str: | |
| return f"Perform TTS in {language} using a {gender} voice." | |