import os import json from functools import lru_cache from pathlib import Path os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1") os.environ.setdefault("HF_HUB_DISABLE_TELEMETRY", "1") import sherpa_onnx from huggingface_hub import hf_hub_download, snapshot_download def get_file(repo_id: str, filename: str, subfolder: str = ".") -> str: return hf_hub_download(repo_id=repo_id, filename=filename, subfolder=subfolder) @lru_cache(maxsize=10) def _get_kokoro(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts: data_dir = "/tmp/espeak-ng-data" clean_id = repo_id.split("|")[0] is_int8 = "int8" in clean_id model_name = "model.int8.onnx" if is_int8 else "model.onnx" model = get_file(repo_id=clean_id, filename=model_name, subfolder=".") tokens = get_file(repo_id=clean_id, filename="tokens.txt", subfolder=".") voices = get_file(repo_id=clean_id, filename="voices.bin", subfolder=".") if "multi-lang" in clean_id: lexicon_en = get_file(repo_id=clean_id, filename="lexicon-us-en.txt", subfolder=".") lexicon_zh = get_file(repo_id=clean_id, filename="lexicon-zh.txt", subfolder=".") lexicon = f"{lexicon_en},{lexicon_zh}" date_zh = get_file(repo_id=clean_id, filename="date-zh.fst", subfolder=".") number_zh = get_file(repo_id=clean_id, filename="number-zh.fst", subfolder=".") phone_zh = get_file(repo_id=clean_id, filename="phone-zh.fst", subfolder=".") rule_fsts = f"{date_zh},{phone_zh},{number_zh}" dict_dir = "/tmp/dict" else: lexicon = "" rule_fsts = "" dict_dir = "" tts_config = sherpa_onnx.OfflineTtsConfig( model=sherpa_onnx.OfflineTtsModelConfig( kokoro=sherpa_onnx.OfflineTtsKokoroModelConfig( model=model, voices=voices, tokens=tokens, data_dir=data_dir, length_scale=1.0 / speed, lexicon=lexicon, dict_dir=dict_dir, ), provider="cpu", debug=False, num_threads=2, ), max_num_sentences=1, rule_fsts=rule_fsts, ) return sherpa_onnx.OfflineTts(tts_config) @lru_cache(maxsize=10) def _get_supertonic(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts: clean_id = repo_id.split("|")[0] dp = get_file(repo_id=clean_id, filename="duration_predictor.int8.onnx", subfolder=".") te = get_file(repo_id=clean_id, filename="text_encoder.int8.onnx", subfolder=".") ve = get_file(repo_id=clean_id, filename="vector_estimator.int8.onnx", subfolder=".") vo = get_file(repo_id=clean_id, filename="vocoder.int8.onnx", subfolder=".") tts_json = get_file(repo_id=clean_id, filename="tts.json", subfolder=".") ui = get_file(repo_id=clean_id, filename="unicode_indexer.bin", subfolder=".") vs = get_file(repo_id=clean_id, filename="voice.bin", subfolder=".") tts_config = sherpa_onnx.OfflineTtsConfig( model=sherpa_onnx.OfflineTtsModelConfig( supertonic=sherpa_onnx.OfflineTtsSupertonicModelConfig( duration_predictor=dp, text_encoder=te, vector_estimator=ve, vocoder=vo, tts_json=tts_json, unicode_indexer=ui, voice_style=vs, ), provider="cpu", debug=False, num_threads=2, ), max_num_sentences=1, ) return sherpa_onnx.OfflineTts(tts_config) @lru_cache(maxsize=10) def _get_vits_piper(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts: data_dir = "/tmp/espeak-ng-data" lexicon = "" rule_fsts = "" clean_id = repo_id.split("|")[0] if "piper" in clean_id: n = len("vits-piper-") name = clean_id.split("/")[1][n:] elif "mimic3" in clean_id: n = len("vits-mimic3-") name = clean_id.split("/")[1][n:] else: name = "model" local_dir = snapshot_download(clean_id) model = f"{local_dir}/{name}.onnx" tokens = f"{local_dir}/tokens.txt" if "vits-piper-zh_CN-chaowen-medium" in clean_id or "vits-piper-zh_CN-xiao_ya-medium" in clean_id: data_dir = "" lexicon = f"{local_dir}/lexicon.txt" rule_fsts = ["phone.fst", "date.fst", "number.fst"] rule_fsts = ",".join(f"{local_dir}/{r}" for r in rule_fsts) tts_config = sherpa_onnx.OfflineTtsConfig( model=sherpa_onnx.OfflineTtsModelConfig( vits=sherpa_onnx.OfflineTtsVitsModelConfig( model=model, lexicon=lexicon, data_dir=data_dir, tokens=tokens, length_scale=1.0 / speed, ), matcha=sherpa_onnx.OfflineTtsMatchaModelConfig(), provider="cpu", debug=False, num_threads=2, ), max_num_sentences=1, rule_fsts=rule_fsts, ) return sherpa_onnx.OfflineTts(tts_config) @lru_cache(maxsize=10) def _get_vits_zh_aishell3(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts: clean_id = repo_id.split("|")[0] model = get_file(repo_id=clean_id, filename="vits-aishell3.onnx", subfolder=".") lexicon = get_file(repo_id=clean_id, filename="lexicon.txt", subfolder=".") tokens = get_file(repo_id=clean_id, filename="tokens.txt", subfolder=".") rule_fsts = ["phone.fst", "date.fst", "number.fst", "new_heteronym.fst"] rule_fsts = ",".join(get_file(repo_id=clean_id, filename=f, subfolder=".") for f in rule_fsts) rule_fars = get_file(repo_id=clean_id, filename="rule.far", subfolder=".") tts_config = sherpa_onnx.OfflineTtsConfig( model=sherpa_onnx.OfflineTtsModelConfig( vits=sherpa_onnx.OfflineTtsVitsModelConfig( model=model, lexicon=lexicon, tokens=tokens, length_scale=1.0 / speed, ), matcha=sherpa_onnx.OfflineTtsMatchaModelConfig(), provider="cpu", debug=False, num_threads=2, ), rule_fsts=rule_fsts, rule_fars=rule_fars, max_num_sentences=1, ) return sherpa_onnx.OfflineTts(tts_config) @lru_cache(maxsize=10) def _get_matcha_zh_en(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts: clean_id = repo_id.split("|")[0] acoustic_model = get_file(repo_id=clean_id, filename="model-steps-3.onnx", subfolder=".") vocoder = get_file(repo_id="csukuangfj/sherpa-onnx-vocoders", filename="vocos-16khz-univ.onnx", subfolder=".") lexicon = get_file(repo_id=clean_id, filename="lexicon.txt", subfolder=".") tokens = get_file(repo_id=clean_id, filename="tokens.txt", subfolder=".") rule_fsts = ["phone-zh.fst", "date-zh.fst", "number-zh.fst"] rule_fsts = ",".join(get_file(repo_id=clean_id, filename=f, subfolder=".") for f in rule_fsts) tts_config = sherpa_onnx.OfflineTtsConfig( model=sherpa_onnx.OfflineTtsModelConfig( vits=sherpa_onnx.OfflineTtsVitsModelConfig(), matcha=sherpa_onnx.OfflineTtsMatchaModelConfig( acoustic_model=acoustic_model, vocoder=vocoder, lexicon=lexicon, tokens=tokens, data_dir="/tmp/espeak-ng-data", length_scale=1.0 / speed, ), provider="cpu", debug=False, num_threads=2, ), rule_fsts=rule_fsts, rule_fars="", max_num_sentences=1, ) return sherpa_onnx.OfflineTts(tts_config) @lru_cache(maxsize=10) def _get_matcha_hf(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts: clean_id = repo_id.split("|")[0] if not Path("/tmp/dict").is_dir(): os.system("cd /tmp; curl -SL -O https://github.com/csukuangfj/cppjieba/releases/download/sherpa-onnx-2024-04-19/dict.tar.bz2; tar xf dict.tar.bz2") acoustic_model = get_file(repo_id=clean_id, filename="model-steps-3.onnx", subfolder=".") vocoder = get_file(repo_id="csukuangfj/sherpa-onnx-hifigan", filename="hifigan_v2.onnx", subfolder=".") lexicon = get_file(repo_id=clean_id, filename="lexicon.txt", subfolder=".") tokens = get_file(repo_id=clean_id, filename="tokens.txt", subfolder=".") rule_fsts = ["phone.fst", "date.fst", "number.fst"] rule_fsts = ",".join(get_file(repo_id=clean_id, filename=f, subfolder=".") for f in rule_fsts) tts_config = sherpa_onnx.OfflineTtsConfig( model=sherpa_onnx.OfflineTtsModelConfig( vits=sherpa_onnx.OfflineTtsVitsModelConfig(), matcha=sherpa_onnx.OfflineTtsMatchaModelConfig( acoustic_model=acoustic_model, vocoder=vocoder, lexicon=lexicon, tokens=tokens, dict_dir="/tmp/dict", length_scale=1.0 / speed, ), provider="cpu", debug=False, num_threads=2, ), rule_fsts=rule_fsts, rule_fars="", max_num_sentences=1, ) return sherpa_onnx.OfflineTts(tts_config) @lru_cache(maxsize=10) def _get_vits_hf(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts: clean_id = repo_id.split("|")[0] if "fanchen" in clean_id or "vits-cantonese-hf-xiaomaiiwn" in clean_id: model = clean_id.split("/")[-1] elif "vits-melo-tts" in clean_id: model = "model" else: model = "model" if "vits-zh-ll" in clean_id or "vits-melo-tts" in clean_id: if not Path("/tmp/dict").is_dir(): os.system("cd /tmp; curl -SL -O https://github.com/csukuangfj/cppjieba/releases/download/sherpa-onnx-2024-04-19/dict.tar.bz2; tar xf dict.tar.bz2") model_file = get_file(repo_id=clean_id, filename=f"{model}.onnx", subfolder=".") lexicon = get_file(repo_id=clean_id, filename="lexicon.txt", subfolder=".") tokens = get_file(repo_id=clean_id, filename="tokens.txt", subfolder=".") if "vits-cantonese-hf-xiaomaiiwn" not in clean_id: rule_fsts = ["phone.fst", "date.fst", "number.fst"] rule_fsts = ",".join(get_file(repo_id=clean_id, filename=f, subfolder=".") for f in rule_fsts) vits_dict_dir = "/tmp/dict" else: rule_fsts = get_file(repo_id=clean_id, filename="rule.fst", subfolder=".") vits_dict_dir = "" tts_config = sherpa_onnx.OfflineTtsConfig( model=sherpa_onnx.OfflineTtsModelConfig( vits=sherpa_onnx.OfflineTtsVitsModelConfig( model=model_file, lexicon=lexicon, tokens=tokens, dict_dir=vits_dict_dir, length_scale=1.0 / speed, ), matcha=sherpa_onnx.OfflineTtsMatchaModelConfig(), provider="cpu", debug=False, num_threads=2, ), rule_fsts=rule_fsts, rule_fars="", max_num_sentences=1, ) return sherpa_onnx.OfflineTts(tts_config) @lru_cache(maxsize=10) def _get_melotts_onnx(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts: clean_id = repo_id.split("|")[0] parts = clean_id.split("/") lang_subdir = parts[2] if len(parts) > 2 else parts[1] model = hf_hub_download(repo_id="MiaoMint/MeloTTS-ONNX", filename=f"onnx_exports/{lang_subdir}/model.onnx") tokens = hf_hub_download(repo_id="MiaoMint/MeloTTS-ONNX", filename=f"onnx_exports/{lang_subdir}/tokens.txt") lexicon_path = "" try: lexicon_path = hf_hub_download(repo_id="MiaoMint/MeloTTS-ONNX", filename=f"onnx_exports/{lang_subdir}/lexicon.txt") except Exception: pass data_dir = "/tmp/espeak-ng-data" dict_dir = "/tmp/dict" if lang_subdir == "zh" else "" if dict_dir and not Path(dict_dir).is_dir(): os.system("cd /tmp; curl -SL -O https://github.com/csukuangfj/cppjieba/releases/download/sherpa-onnx-2024-04-19/dict.tar.bz2; tar xf dict.tar.bz2") tts_config = sherpa_onnx.OfflineTtsConfig( model=sherpa_onnx.OfflineTtsModelConfig( vits=sherpa_onnx.OfflineTtsVitsModelConfig( model=model, lexicon=lexicon_path, data_dir=data_dir, tokens=tokens, dict_dir=dict_dir, length_scale=1.0 / speed, ), matcha=sherpa_onnx.OfflineTtsMatchaModelConfig(), provider="cpu", debug=False, num_threads=2, ), max_num_sentences=1, ) return sherpa_onnx.OfflineTts(tts_config) chinese_models = { "csukuangfj2/vits-piper-zh_CN-chaowen-medium|1 speaker": _get_vits_piper, "csukuangfj2/vits-piper-zh_CN-xiao_ya-medium|1 speaker": _get_vits_piper, "csukuangfj/matcha-icefall-zh-baker|1 speaker": _get_matcha_hf, "csukuangfj/vits-zh-hf-fanchen-wnj|1 speaker": _get_vits_hf, "csukuangfj/vits-zh-hf-fanchen-C|187 speakers": _get_vits_hf, "csukuangfj/sherpa-onnx-vits-zh-ll|5 speakers": _get_vits_hf, "csukuangfj/vits-zh-aishell3|174 speakers": _get_vits_zh_aishell3, "csukuangfj/vits-piper-zh_CN-huayan-medium|1 speaker": _get_vits_piper, "MiaoMint/MeloTTS-ONNX/zh|1 speaker": _get_melotts_onnx, } chinese_english_models = { "csukuangfj/matcha-icefall-zh-en|1": _get_matcha_zh_en, "csukuangfj/kokoro-multi-lang-v1_1|103 speakers": _get_kokoro, "csukuangfj/kokoro-int8-multi-lang-v1_1|103 speakers": _get_kokoro, "csukuangfj/kokoro-multi-lang-v1_0|53 speakers": _get_kokoro, "csukuangfj/kokoro-int8-multi-lang-v1_0|53 speakers": _get_kokoro, "csukuangfj2/sherpa-onnx-supertonic-3-tts-int8-2026-05-11|10 speakers": _get_supertonic, "csukuangfj/vits-melo-tts-zh_en|1": _get_vits_hf, "MiaoMint/MeloTTS-ONNX/zh|1 speaker": _get_melotts_onnx, } cantonese_models = { "csukuangfj/vits-cantonese-hf-xiaomaiiwn|1 speaker": _get_vits_hf, } english_models = { "csukuangfj/kokoro-en-v0_19|11 speakers": _get_kokoro, "csukuangfj2/sherpa-onnx-supertonic-3-tts-int8-2026-05-11|10 speakers": _get_supertonic, "MiaoMint/MeloTTS-ONNX/en_newest|1 speaker": _get_melotts_onnx, "csukuangfj/vits-piper-en_US-lessac-high|1 speaker": _get_vits_piper, "csukuangfj/vits-piper-en_US-ryan-high|1 speaker": _get_vits_piper, "csukuangfj/vits-piper-en_GB-alan-medium|1 speaker": _get_vits_piper, "csukuangfj/vits-piper-en_GB-jenny_dioco-medium|1 speaker": _get_vits_piper, "csukuangfj/vits-piper-en_GB-vctk-medium|109 speakers": _get_vits_piper, } language_to_models = { "中文（普通话）": list(chinese_models.keys()), "中英双语": list(chinese_english_models.keys()), "粤语": list(cantonese_models.keys()), "英语": list(english_models.keys()), } all_model_dicts = { **chinese_models, **chinese_english_models, **cantonese_models, **english_models, } @lru_cache(maxsize=32) def get_speaker_map(repo_id: str) -> dict: clean_id = repo_id.split("|")[0] if "piper" in clean_id: try: local_dir = snapshot_download(clean_id) for fn in os.listdir(local_dir): if fn.endswith(".onnx.json"): with open(os.path.join(local_dir, fn)) as f: data = json.load(f) sid_map = data.get("speaker_id_map", {}) return {v: k for k, v in sid_map.items()} except Exception: pass if "vits-zh-hf-fanchen-C" in clean_id: try: f = hf_hub_download(repo_id=clean_id, filename="G_C.json") with open(f) as fp: data = json.load(fp) speakers = data.get("speakers", []) return {i: s for i, s in enumerate(speakers)} except Exception: pass if "vits-zh-aishell3" in clean_id: return {} if "kokoro" in clean_id: kokoro_en = [ (0, "美式女声-af"), (1, "美式女声-bella"), (2, "美式女声-nicole"), (3, "美式女声-sarah"), (4, "美式女声-sky"), (5, "美式男声-adam"), (6, "美式男声-michael"), (7, "英式女声-emma"), (8, "英式女声-isabella"), (9, "英式男声-george"), (10, "英式男声-lewis"), ] kokoro_multi_v10 = [ (0, "美式女声-alloy"), (1, "美式女声-aoede"), (2, "美式女声-bella"), (3, "美式女声-heart"), (4, "美式女声-jessica"), (5, "美式女声-kore"), (6, "美式女声-nicole"), (7, "美式女声-nova"), (8, "美式女声-river"), (9, "美式女声-sarah"), (10, "美式女声-sky"), (11, "美式男声-adam"), (12, "美式男声-echo"), (13, "美式男声-eric"), (14, "美式男声-fenrir"), (15, "美式男声-liam"), (16, "美式男声-michael"), (17, "美式男声-onyx"), (18, "美式男声-puck"), (19, "美式男声-santa"), (20, "英式女声-alice"), (21, "英式女声-emma"), (22, "英式女声-isabella"), (23, "英式女声-lily"), (24, "英式男声-daniel"), (25, "英式男声-fable"), (26, "英式男声-george"), (27, "英式男声-lewis"), (28, "英语女声-dora"), (29, "英语男声-alex"), (30, "法语女声-siwis"), (31, "印地语女声-alpha"), (32, "印地语女声-beta"), (33, "印地语男声-omega"), (34, "印地语男声-psi"), (35, "意语女声-sara"), (36, "意语男声-nicola"), (37, "日语女声-alpha"), (38, "日语女声-gongitsune"), (39, "日语女声-nezumi"), (40, "日语女声-tebukuro"), (41, "日语男声-kumo"), (42, "葡语女声-dora"), (43, "葡语男声-alex"), (44, "葡语男声-santa"), (45, "中文女声-小北"), (46, "中文女声-小妮"), (47, "中文女声-小小"), (48, "中文女声-小艺"), (49, "中文男声-云剑"), (50, "中文男声-云希"), (51, "中文男声-云夏"), (52, "中文男声-云扬"), ] kokoro_multi_v11 = [ (0, "美式女声-maple"), (1, "美式女声-sol"), (2, "英式女声-vale"), (3, "中文女声-001"), (4, "中文女声-002"), (5, "中文女声-003"), (6, "中文女声-004"), (7, "中文女声-005"), (8, "中文女声-006"), (9, "中文女声-007"), (10, "中文女声-008"), (11, "中文女声-017"), (12, "中文女声-018"), (13, "中文女声-019"), (14, "中文女声-021"), (15, "中文女声-022"), (16, "中文女声-023"), (17, "中文女声-024"), (18, "中文女声-026"), (19, "中文女声-027"), (20, "中文女声-028"), (21, "中文女声-032"), (22, "中文女声-036"), (23, "中文女声-038"), (24, "中文女声-039"), (25, "中文女声-040"), (26, "中文女声-042"), (27, "中文女声-043"), (28, "中文女声-044"), (29, "中文女声-046"), (30, "中文女声-047"), (31, "中文女声-048"), (32, "中文女声-049"), (33, "中文女声-051"), (34, "中文女声-059"), (35, "中文女声-060"), (36, "中文女声-067"), (37, "中文女声-070"), (38, "中文女声-071"), (39, "中文女声-072"), (40, "中文女声-073"), (41, "中文女声-074"), (42, "中文女声-075"), (43, "中文女声-076"), (44, "中文女声-077"), (45, "中文女声-078"), (46, "中文女声-079"), (47, "中文女声-083"), (48, "中文女声-084"), (49, "中文女声-085"), (50, "中文女声-086"), (51, "中文女声-087"), (52, "中文女声-088"), (53, "中文女声-090"), (54, "中文女声-092"), (55, "中文女声-093"), (56, "中文女声-094"), (57, "中文女声-099"), (58, "中文男声-009"), (59, "中文男声-010"), (60, "中文男声-011"), (61, "中文男声-012"), (62, "中文男声-013"), (63, "中文男声-014"), (64, "中文男声-015"), (65, "中文男声-016"), (66, "中文男声-020"), (67, "中文男声-025"), (68, "中文男声-029"), (69, "中文男声-030"), (70, "中文男声-031"), (71, "中文男声-033"), (72, "中文男声-034"), (73, "中文男声-035"), (74, "中文男声-037"), (75, "中文男声-041"), (76, "中文男声-045"), (77, "中文男声-050"), (78, "中文男声-052"), (79, "中文男声-053"), (80, "中文男声-054"), (81, "中文男声-055"), (82, "中文男声-056"), (83, "中文男声-057"), (84, "中文男声-058"), (85, "中文男声-061"), (86, "中文男声-062"), (87, "中文男声-063"), (88, "中文男声-064"), (89, "中文男声-065"), (90, "中文男声-066"), (91, "中文男声-068"), (92, "中文男声-069"), (93, "中文男声-080"), (94, "中文男声-081"), (95, "中文男声-082"), (96, "中文男声-089"), (97, "中文男声-091"), (98, "中文男声-095"), (99, "中文男声-096"), (100, "中文男声-097"), (101, "中文男声-098"), (102, "中文男声-100"), ] kokoro_maps = { "csukuangfj/kokoro-en-v0_19": kokoro_en, "csukuangfj/kokoro-multi-lang-v1_1": kokoro_multi_v11, "csukuangfj/kokoro-multi-lang-v1_0": kokoro_multi_v10, "csukuangfj/kokoro-int8-multi-lang-v1_1": kokoro_multi_v11, "csukuangfj/kokoro-int8-multi-lang-v1_0": kokoro_multi_v10, } if clean_id in kokoro_maps: return {sid: name for sid, name in kokoro_maps[clean_id]} if "supertonic" in clean_id: return { 0: "男声-M1（活泼自信）", 1: "男声-M2（深沉稳重）", 2: "男声-M3（专业权威）", 3: "男声-M4（柔和亲切）", 4: "男声-M5（温暖舒缓）", 5: "女声-F1（沉稳从容）", 6: "女声-F2（明快活泼）", 7: "女声-F3（清晰专业）", 8: "女声-F4（干练自信）", 9: "女声-F5（温柔平和）", } if "vits-zh-ll" in clean_id: return {0: "女声0", 1: "女声1", 2: "男声0", 3: "男声1", 4: "男声2"} return {} @lru_cache(maxsize=10) def get_pretrained_model(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts: if repo_id in all_model_dicts: return all_model_dicts[repo_id](repo_id, speed) raise ValueError(f"不支持的模型: {repo_id}")