Merge pull request #436 from FunAudioLLM/dev/lyuxiang.lx

Dev/lyuxiang.lx

README.md

@@ -22,12 +22,9 @@ For `SenseVoice`, visit [SenseVoice repo](https://github.com/FunAudioLLM/SenseVo
     - [ ] 25hz cosyvoice base model
     - [ ] 25hz cosyvoice voice conversion model
 
-- [ ] 2024/10
-
-    - [ ] 50hz llama based llm model which supports lora finetune
-
 - [ ] TBD
 
+    - [ ] 25hz llama based llm model which supports lora finetune
     - [ ] Support more instruction mode
     - [ ] Voice conversion
     - [ ] Music generation
@@ -74,6 +71,7 @@ If you are expert in this field, and you are only interested in training your ow
 # SDK模型下载
 from modelscope import snapshot_download
 snapshot_download('iic/CosyVoice-300M', local_dir='pretrained_models/CosyVoice-300M')
+snapshot_download('iic/CosyVoice-300M-25Hz', local_dir='pretrained_models/CosyVoice-300M-25Hz')
 snapshot_download('iic/CosyVoice-300M-SFT', local_dir='pretrained_models/CosyVoice-300M-SFT')
 snapshot_download('iic/CosyVoice-300M-Instruct', local_dir='pretrained_models/CosyVoice-300M-Instruct')
 snapshot_download('iic/CosyVoice-ttsfrd', local_dir='pretrained_models/CosyVoice-ttsfrd')
@@ -83,6 +81,7 @@ snapshot_download('iic/CosyVoice-ttsfrd', local_dir='pretrained_models/CosyVoice
 # git模型下载，请确保已安装git lfs
 mkdir -p pretrained_models
 git clone https://www.modelscope.cn/iic/CosyVoice-300M.git pretrained_models/CosyVoice-300M
+git clone https://www.modelscope.cn/iic/CosyVoice-300M-25Hz.git pretrained_models/CosyVoice-300M-25Hz
 git clone https://www.modelscope.cn/iic/CosyVoice-300M-SFT.git pretrained_models/CosyVoice-300M-SFT
 git clone https://www.modelscope.cn/iic/CosyVoice-300M-Instruct.git pretrained_models/CosyVoice-300M-Instruct
 git clone https://www.modelscope.cn/iic/CosyVoice-ttsfrd.git pretrained_models/CosyVoice-ttsfrd
@@ -121,7 +120,7 @@ print(cosyvoice.list_avaliable_spks())
 for i, j in enumerate(cosyvoice.inference_sft('你好，我是通义生成式语音大模型，请问有什么可以帮您的吗？', '中文女', stream=False)):
     torchaudio.save('sft_{}.wav'.format(i), j['tts_speech'], 22050)
 
-cosyvoice = CosyVoice('pretrained_models/CosyVoice-300M')
+cosyvoice = CosyVoice('pretrained_models/CosyVoice-300M-25Hz') # or change to pretrained_models/CosyVoice-300M for 50Hz inference
 # zero_shot usage, <|zh|><|en|><|jp|><|yue|><|ko|> for Chinese/English/Japanese/Cantonese/Korean
 prompt_speech_16k = load_wav('zero_shot_prompt.wav', 16000)
 for i, j in enumerate(cosyvoice.inference_zero_shot('收到好友从远方寄来的生日礼物，那份意外的惊喜与深深的祝福让我心中充满了甜蜜的快乐，笑容如花儿般绽放。', '希望你以后能够做的比我还好呦。', prompt_speech_16k, stream=False)):
@@ -130,6 +129,11 @@ for i, j in enumerate(cosyvoice.inference_zero_shot('收到好友从远方寄来
 prompt_speech_16k = load_wav('cross_lingual_prompt.wav', 16000)
 for i, j in enumerate(cosyvoice.inference_cross_lingual('<|en|>And then later on, fully acquiring that company. So keeping management in line, interest in line with the asset that\'s coming into the family is a reason why sometimes we don\'t buy the whole thing.', prompt_speech_16k, stream=False)):
     torchaudio.save('cross_lingual_{}.wav'.format(i), j['tts_speech'], 22050)
+# vc usage
+prompt_speech_16k = load_wav('zero_shot_prompt.wav', 16000)
+source_speech_16k = load_wav('cross_lingual_prompt.wav', 16000)
+for i, j in enumerate(cosyvoice.inference_vc(source_speech_16k, prompt_speech_16k, stream=False)):
+    torchaudio.save('vc_{}.wav'.format(i), j['tts_speech'], 22050)
 
 cosyvoice = CosyVoice('pretrained_models/CosyVoice-300M-Instruct')
 # instruct usage, support <laughter></laughter><strong></strong>[laughter][breath]

cosyvoice/cli/cosyvoice.py

@@ -58,7 +58,7 @@ class CosyVoice:
             model_input = self.frontend.frontend_sft(i, spk_id)
             start_time = time.time()
             logging.info('synthesis text {}'.format(i))
-            for model_output in self.model.inference(**model_input, stream=stream, speed=speed):
+            for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
                 speech_len = model_output['tts_speech'].shape[1] / 22050
                 logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
                 yield model_output
@@ -70,7 +70,7 @@ class CosyVoice:
             model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k)
             start_time = time.time()
             logging.info('synthesis text {}'.format(i))
-            for model_output in self.model.inference(**model_input, stream=stream, speed=speed):
+            for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
                 speech_len = model_output['tts_speech'].shape[1] / 22050
                 logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
                 yield model_output
@@ -83,7 +83,7 @@ class CosyVoice:
             model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k)
             start_time = time.time()
             logging.info('synthesis text {}'.format(i))
-            for model_output in self.model.inference(**model_input, stream=stream, speed=speed):
+            for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
                 speech_len = model_output['tts_speech'].shape[1] / 22050
                 logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
                 yield model_output
@@ -97,8 +97,17 @@ class CosyVoice:
             model_input = self.frontend.frontend_instruct(i, spk_id, instruct_text)
             start_time = time.time()
             logging.info('synthesis text {}'.format(i))
-            for model_output in self.model.inference(**model_input, stream=stream, speed=speed):
+            for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
                 speech_len = model_output['tts_speech'].shape[1] / 22050
                 logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
                 yield model_output
                 start_time = time.time()
+
+    def inference_vc(self, source_speech_16k, prompt_speech_16k, stream=False, speed=1.0):
+        model_input = self.frontend.frontend_vc(source_speech_16k, prompt_speech_16k)
+        start_time = time.time()
+        for model_output in self.model.vc(**model_input, stream=stream, speed=speed):
+            speech_len = model_output['tts_speech'].shape[1] / 22050
+            logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
+            yield model_output
+            start_time = time.time()

cosyvoice/cli/frontend.py

@@ -55,6 +55,8 @@ class CosyVoiceFrontEnd:
                                                                                 "CPUExecutionProvider"])
         if os.path.exists(spk2info):
             self.spk2info = torch.load(spk2info, map_location=self.device)
+        else:
+            self.spk2info = {}
         self.instruct = instruct
         self.allowed_special = allowed_special
         self.inflect_parser = inflect.engine()
@@ -172,3 +174,15 @@ class CosyVoiceFrontEnd:
         model_input['prompt_text'] = instruct_text_token
         model_input['prompt_text_len'] = instruct_text_token_len
         return model_input
+
+    def frontend_vc(self, source_speech_16k, prompt_speech_16k):
+        prompt_speech_token, prompt_speech_token_len = self._extract_speech_token(prompt_speech_16k)
+        prompt_speech_22050 = torchaudio.transforms.Resample(orig_freq=16000, new_freq=22050)(prompt_speech_16k)
+        prompt_speech_feat, prompt_speech_feat_len = self._extract_speech_feat(prompt_speech_22050)
+        embedding = self._extract_spk_embedding(prompt_speech_16k)
+        source_speech_token, source_speech_token_len = self._extract_speech_token(source_speech_16k)
+        model_input = {'source_speech_token': source_speech_token, 'source_speech_token_len': source_speech_token_len,
+                       'flow_prompt_speech_token': prompt_speech_token, 'flow_prompt_speech_token_len': prompt_speech_token_len,
+                       'prompt_speech_feat': prompt_speech_feat, 'prompt_speech_feat_len': prompt_speech_feat_len,
+                       'flow_embedding': embedding}
+        return model_input

cosyvoice/cli/model.py

@@ -35,7 +35,7 @@ class CosyVoiceModel:
         self.token_max_hop_len = 200
         self.token_overlap_len = 20
         # mel fade in out
-        self.mel_overlap_len = 34
+        self.mel_overlap_len = int(self.token_overlap_len / self.flow.input_frame_rate * 22050 / 256)
         self.mel_window = np.hamming(2 * self.mel_overlap_len)
         # hift cache
         self.mel_cache_len = 20
@@ -63,11 +63,11 @@ class CosyVoiceModel:
         self.hift.to(self.device).eval()
 
     def load_jit(self, llm_text_encoder_model, llm_llm_model, flow_encoder_model):
-        llm_text_encoder = torch.jit.load(llm_text_encoder_model)
+        llm_text_encoder = torch.jit.load(llm_text_encoder_model, map_location=self.device)
         self.llm.text_encoder = llm_text_encoder
-        llm_llm = torch.jit.load(llm_llm_model)
+        llm_llm = torch.jit.load(llm_llm_model, map_location=self.device)
         self.llm.llm = llm_llm
-        flow_encoder = torch.jit.load(flow_encoder_model)
+        flow_encoder = torch.jit.load(flow_encoder_model, map_location=self.device)
         self.flow.encoder = flow_encoder
 
     def load_onnx(self, flow_decoder_estimator_model):
@@ -131,11 +131,11 @@ class CosyVoiceModel:
                 tts_speech = fade_in_out(tts_speech, self.hift_cache_dict[uuid]['speech'], self.speech_window)
         return tts_speech
 
-    def inference(self, text, flow_embedding, llm_embedding=torch.zeros(0, 192),
-                  prompt_text=torch.zeros(1, 0, dtype=torch.int32),
-                  llm_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32),
-                  flow_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32),
-                  prompt_speech_feat=torch.zeros(1, 0, 80), stream=False, speed=1.0, **kwargs):
+    def tts(self, text, flow_embedding, llm_embedding=torch.zeros(0, 192),
+            prompt_text=torch.zeros(1, 0, dtype=torch.int32),
+            llm_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32),
+            flow_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32),
+            prompt_speech_feat=torch.zeros(1, 0, 80), stream=False, speed=1.0, **kwargs):
         # this_uuid is used to track variables related to this inference thread
         this_uuid = str(uuid.uuid1())
         with self.lock:
@@ -148,7 +148,8 @@ class CosyVoiceModel:
             while True:
                 time.sleep(0.1)
                 if len(self.tts_speech_token_dict[this_uuid]) >= token_hop_len + self.token_overlap_len:
-                    this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid][:token_hop_len + self.token_overlap_len], dim=1)
+                    this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid][:token_hop_len + self.token_overlap_len]) \
+                        .unsqueeze(dim=0)
                     this_tts_speech = self.token2wav(token=this_tts_speech_token,
                                                      prompt_token=flow_prompt_speech_token,
                                                      prompt_feat=prompt_speech_feat,
@@ -164,7 +165,7 @@ class CosyVoiceModel:
                     break
             p.join()
             # deal with remain tokens, make sure inference remain token len equals token_hop_len when cache_speech is not None
-            this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid], dim=1)
+            this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid]).unsqueeze(dim=0)
             this_tts_speech = self.token2wav(token=this_tts_speech_token,
                                              prompt_token=flow_prompt_speech_token,
                                              prompt_feat=prompt_speech_feat,
@@ -175,7 +176,58 @@ class CosyVoiceModel:
         else:
             # deal with all tokens
             p.join()
-            this_tts_speech_token = torch.concat(self.tts_speech_token_dict[this_uuid], dim=1)
+            this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid]).unsqueeze(dim=0)
+            this_tts_speech = self.token2wav(token=this_tts_speech_token,
+                                             prompt_token=flow_prompt_speech_token,
+                                             prompt_feat=prompt_speech_feat,
+                                             embedding=flow_embedding,
+                                             uuid=this_uuid,
+                                             finalize=True,
+                                             speed=speed)
+            yield {'tts_speech': this_tts_speech.cpu()}
+        with self.lock:
+            self.tts_speech_token_dict.pop(this_uuid)
+            self.llm_end_dict.pop(this_uuid)
+            self.mel_overlap_dict.pop(this_uuid)
+            self.hift_cache_dict.pop(this_uuid)
+
+    def vc(self, source_speech_token, flow_prompt_speech_token, prompt_speech_feat, flow_embedding, stream=False, speed=1.0, **kwargs):
+        # this_uuid is used to track variables related to this inference thread
+        this_uuid = str(uuid.uuid1())
+        with self.lock:
+            self.tts_speech_token_dict[this_uuid], self.llm_end_dict[this_uuid] = source_speech_token.flatten().tolist(), True
+            self.mel_overlap_dict[this_uuid], self.hift_cache_dict[this_uuid] = None, None
+        if stream is True:
+            token_hop_len = self.token_min_hop_len
+            while True:
+                if len(self.tts_speech_token_dict[this_uuid]) >= token_hop_len + self.token_overlap_len:
+                    this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid][:token_hop_len + self.token_overlap_len]) \
+                        .unsqueeze(dim=0)
+                    this_tts_speech = self.token2wav(token=this_tts_speech_token,
+                                                     prompt_token=flow_prompt_speech_token,
+                                                     prompt_feat=prompt_speech_feat,
+                                                     embedding=flow_embedding,
+                                                     uuid=this_uuid,
+                                                     finalize=False)
+                    yield {'tts_speech': this_tts_speech.cpu()}
+                    with self.lock:
+                        self.tts_speech_token_dict[this_uuid] = self.tts_speech_token_dict[this_uuid][token_hop_len:]
+                    # increase token_hop_len for better speech quality
+                    token_hop_len = min(self.token_max_hop_len, int(token_hop_len * self.stream_scale_factor))
+                if self.llm_end_dict[this_uuid] is True and len(self.tts_speech_token_dict[this_uuid]) < token_hop_len + self.token_overlap_len:
+                    break
+            # deal with remain tokens, make sure inference remain token len equals token_hop_len when cache_speech is not None
+            this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid], dim=1).unsqueeze(dim=0)
+            this_tts_speech = self.token2wav(token=this_tts_speech_token,
+                                             prompt_token=flow_prompt_speech_token,
+                                             prompt_feat=prompt_speech_feat,
+                                             embedding=flow_embedding,
+                                             uuid=this_uuid,
+                                             finalize=True)
+            yield {'tts_speech': this_tts_speech.cpu()}
+        else:
+            # deal with all tokens
+            this_tts_speech_token = torch.tensor(self.tts_speech_token_dict[this_uuid]).unsqueeze(dim=0)
             this_tts_speech = self.token2wav(token=this_tts_speech_token,
                                              prompt_token=flow_prompt_speech_token,
                                              prompt_feat=prompt_speech_feat,

cosyvoice/flow/flow.py

@@ -124,15 +124,14 @@ class MaskedDiffWithXvec(torch.nn.Module):
         # text encode
         h, h_lengths = self.encoder(token, token_len)
         h = self.encoder_proj(h)
-        mel_len1, mel_len2 = prompt_feat.shape[1], int(token_len2 / 50 * 22050 / 256)
-        h, h_lengths = self.length_regulator.inference(h[:, :token_len1], h[:, token_len1:], mel_len1, mel_len2)
+        mel_len1, mel_len2 = prompt_feat.shape[1], int(token_len2 / self.input_frame_rate * 22050 / 256)
+        h, h_lengths = self.length_regulator.inference(h[:, :token_len1], h[:, token_len1:], mel_len1, mel_len2, self.input_frame_rate)
 
         # get conditions
         conds = torch.zeros([1, mel_len1 + mel_len2, self.output_size], device=token.device)
         conds[:, :mel_len1] = prompt_feat
         conds = conds.transpose(1, 2)
 
-        # mask = (~make_pad_mask(feat_len)).to(h)
         mask = (~make_pad_mask(torch.tensor([mel_len1 + mel_len2]))).to(h)
         feat = self.decoder(
             mu=h.transpose(1, 2).contiguous(),

cosyvoice/flow/length_regulator.py

@@ -49,13 +49,14 @@ class InterpolateRegulator(nn.Module):
         olens = ylens
         return out * mask, olens
 
-    def inference(self, x1, x2, mel_len1, mel_len2):
+    def inference(self, x1, x2, mel_len1, mel_len2, input_frame_rate=50):
         # in inference mode, interploate prompt token and token(head/mid/tail) seprately, so we can get a clear separation point of mel
         # x in (B, T, D)
         if x2.shape[1] > 40:
-            x2_head = F.interpolate(x2[:, :20].transpose(1, 2).contiguous(), size=34, mode='linear')
-            x2_mid = F.interpolate(x2[:, 20:-20].transpose(1, 2).contiguous(), size=mel_len2 - 34 * 2, mode='linear')
-            x2_tail = F.interpolate(x2[:, -20:].transpose(1, 2).contiguous(), size=34, mode='linear')
+            x2_head = F.interpolate(x2[:, :20].transpose(1, 2).contiguous(), size=int(20 / input_frame_rate * 22050 / 256), mode='linear')
+            x2_mid = F.interpolate(x2[:, 20:-20].transpose(1, 2).contiguous(), size=mel_len2 - int(20 / input_frame_rate * 22050 / 256) * 2,
+                                   mode='linear')
+            x2_tail = F.interpolate(x2[:, -20:].transpose(1, 2).contiguous(), size=int(20 / input_frame_rate * 22050 / 256), mode='linear')
             x2 = torch.concat([x2_head, x2_mid, x2_tail], dim=2)
         else:
             x2 = F.interpolate(x2.transpose(1, 2).contiguous(), size=mel_len2, mode='linear')

cosyvoice/llm/llm.py

@@ -206,7 +206,7 @@ class TransformerLM(torch.nn.Module):
             if top_ids == self.speech_token_size:
                 break
             # in stream mode, yield token one by one
-            yield torch.tensor([[top_ids]], dtype=torch.int64, device=device)
+            yield top_ids
             out_tokens.append(top_ids)
             offset += lm_input.size(1)
             lm_input = self.speech_embedding.weight[top_ids].reshape(1, 1, -1)

cosyvoice/tokenizer/tokenizer.py

@@ -0,0 +1,236 @@
+import base64
+import os
+from functools import lru_cache
+from typing import Optional
+from whisper.tokenizer import Tokenizer
+
+import tiktoken
+
+LANGUAGES = {
+    "en": "english",
+    "zh": "chinese",
+    "de": "german",
+    "es": "spanish",
+    "ru": "russian",
+    "ko": "korean",
+    "fr": "french",
+    "ja": "japanese",
+    "pt": "portuguese",
+    "tr": "turkish",
+    "pl": "polish",
+    "ca": "catalan",
+    "nl": "dutch",
+    "ar": "arabic",
+    "sv": "swedish",
+    "it": "italian",
+    "id": "indonesian",
+    "hi": "hindi",
+    "fi": "finnish",
+    "vi": "vietnamese",
+    "he": "hebrew",
+    "uk": "ukrainian",
+    "el": "greek",
+    "ms": "malay",
+    "cs": "czech",
+    "ro": "romanian",
+    "da": "danish",
+    "hu": "hungarian",
+    "ta": "tamil",
+    "no": "norwegian",
+    "th": "thai",
+    "ur": "urdu",
+    "hr": "croatian",
+    "bg": "bulgarian",
+    "lt": "lithuanian",
+    "la": "latin",
+    "mi": "maori",
+    "ml": "malayalam",
+    "cy": "welsh",
+    "sk": "slovak",
+    "te": "telugu",
+    "fa": "persian",
+    "lv": "latvian",
+    "bn": "bengali",
+    "sr": "serbian",
+    "az": "azerbaijani",
+    "sl": "slovenian",
+    "kn": "kannada",
+    "et": "estonian",
+    "mk": "macedonian",
+    "br": "breton",
+    "eu": "basque",
+    "is": "icelandic",
+    "hy": "armenian",
+    "ne": "nepali",
+    "mn": "mongolian",
+    "bs": "bosnian",
+    "kk": "kazakh",
+    "sq": "albanian",
+    "sw": "swahili",
+    "gl": "galician",
+    "mr": "marathi",
+    "pa": "punjabi",
+    "si": "sinhala",
+    "km": "khmer",
+    "sn": "shona",
+    "yo": "yoruba",
+    "so": "somali",
+    "af": "afrikaans",
+    "oc": "occitan",
+    "ka": "georgian",
+    "be": "belarusian",
+    "tg": "tajik",
+    "sd": "sindhi",
+    "gu": "gujarati",
+    "am": "amharic",
+    "yi": "yiddish",
+    "lo": "lao",
+    "uz": "uzbek",
+    "fo": "faroese",
+    "ht": "haitian creole",
+    "ps": "pashto",
+    "tk": "turkmen",
+    "nn": "nynorsk",
+    "mt": "maltese",
+    "sa": "sanskrit",
+    "lb": "luxembourgish",
+    "my": "myanmar",
+    "bo": "tibetan",
+    "tl": "tagalog",
+    "mg": "malagasy",
+    "as": "assamese",
+    "tt": "tatar",
+    "haw": "hawaiian",
+    "ln": "lingala",
+    "ha": "hausa",
+    "ba": "bashkir",
+    "jw": "javanese",
+    "su": "sundanese",
+    "yue": "cantonese",
+    "minnan": "minnan",
+    "wuyu": "wuyu",
+    "dialect": "dialect",
+    "zh/en": "zh/en",
+    "en/zh": "en/zh",
+}
+
+# language code lookup by name, with a few language aliases
+TO_LANGUAGE_CODE = {
+    **{language: code for code, language in LANGUAGES.items()},
+    "burmese": "my",
+    "valencian": "ca",
+    "flemish": "nl",
+    "haitian": "ht",
+    "letzeburgesch": "lb",
+    "pushto": "ps",
+    "panjabi": "pa",
+    "moldavian": "ro",
+    "moldovan": "ro",
+    "sinhalese": "si",
+    "castilian": "es",
+    "mandarin": "zh",
+}
+
+AUDIO_EVENT = {
+    "ASR": "ASR",
+    "AED": "AED",
+    "SER": "SER",
+    "Speech": "Speech",
+    "/Speech": "/Speech",
+    "BGM": "BGM",
+    "/BGM": "/BGM",
+    "Laughter": "Laughter",
+    "/Laughter": "/Laughter",
+    "Applause": "Applause",
+    "/Applause": "/Applause",
+}
+
+EMOTION = {
+    "HAPPY": "HAPPY",
+    "SAD": "SAD",
+    "ANGRY": "ANGRY",
+    "NEUTRAL": "NEUTRAL",
+}
+
+TTS_Vocal_Token = {
+    "TTS/B": "TTS/B",
+    "TTS/O": "TTS/O",
+    "TTS/Q": "TTS/Q",
+    "TTS/A": "TTS/A",
+    "TTS/CO": "TTS/CO",
+    "TTS/CL": "TTS/CL",
+    "TTS/H": "TTS/H",
+    **{f"TTS/SP{i:02d}": f"TTS/SP{i:02d}" for i in range(1, 14)}
+}
+
+
+@lru_cache(maxsize=None)
+def get_encoding(name: str = "gpt2", num_languages: int = 99):
+    vocab_path = os.path.join(os.path.dirname(__file__), "assets", f"{name}.tiktoken")
+    ranks = {
+        base64.b64decode(token): int(rank)
+        for token, rank in (line.split() for line in open(vocab_path) if line)
+    }
+    n_vocab = len(ranks)
+    special_tokens = {}
+
+    specials = [
+        "<|endoftext|>",
+        "<|startoftranscript|>",
+        *[f"<|{lang}|>" for lang in list(LANGUAGES.keys())[:num_languages]],
+        *[f"<|{audio_event}|>" for audio_event in list(AUDIO_EVENT.keys())],
+        *[f"<|{emotion}|>" for emotion in list(EMOTION.keys())],
+        "<|translate|>",
+        "<|transcribe|>",
+        "<|startoflm|>",
+        "<|startofprev|>",
+        "<|nospeech|>",
+        "<|notimestamps|>",
+        *[f"<|SPECIAL_TOKEN_{i}|>" for i in range(1, 31)],        # register special tokens for ASR
+        *[f"<|{tts}|>" for tts in list(TTS_Vocal_Token.keys())],  # register special tokens for TTS
+        *[f"<|{i * 0.02:.2f}|>" for i in range(1501)],
+    ]
+
+    for token in specials:
+        special_tokens[token] = n_vocab
+        n_vocab += 1
+
+    return tiktoken.Encoding(
+        name=os.path.basename(vocab_path),
+        explicit_n_vocab=n_vocab,
+        pat_str=r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""",
+        mergeable_ranks=ranks,
+        special_tokens=special_tokens,
+    )
+
+
+@lru_cache(maxsize=None)
+def get_tokenizer(
+    multilingual: bool,
+    *,
+    num_languages: int = 99,
+    language: Optional[str] = None,
+    task: Optional[str] = None,  # Literal["transcribe", "translate", None]
+) -> Tokenizer:
+    if language is not None:
+        language = language.lower()
+        if language not in LANGUAGES:
+            if language in TO_LANGUAGE_CODE:
+                language = TO_LANGUAGE_CODE[language]
+            else:
+                raise ValueError(f"Unsupported language: {language}")
+
+    if multilingual:
+        encoding_name = "multilingual_zh_ja_yue_char_del"
+        language = language or "en"
+        task = task or "transcribe"
+    else:
+        encoding_name = "gpt2"
+        language = None
+        task = None
+
+    encoding = get_encoding(name=encoding_name, num_languages=num_languages)
+
+    return Tokenizer(
+        encoding=encoding, num_languages=num_languages, language=language, task=task
+    )

cosyvoice/utils/common.py

@@ -15,8 +15,10 @@
 # Modified from ESPnet(https://github.com/espnet/espnet)
 """Unility functions for Transformer."""
 
+import random
 from typing import List
 
+import numpy as np
 import torch
 
 IGNORE_ID = -1
@@ -142,3 +144,10 @@ def fade_in_out(fade_in_mel, fade_out_mel, window):
     fade_in_mel[..., :mel_overlap_len] = fade_in_mel[..., :mel_overlap_len] * window[:mel_overlap_len] + \
         fade_out_mel[..., -mel_overlap_len:] * window[mel_overlap_len:]
     return fade_in_mel.to(device)
+
+
+def set_all_random_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)

webui.py

@@ -24,6 +24,7 @@ ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
 sys.path.append('{}/third_party/Matcha-TTS'.format(ROOT_DIR))
 from cosyvoice.cli.cosyvoice import CosyVoice
 from cosyvoice.utils.file_utils import load_wav, logging
+from cosyvoice.utils.common import set_all_random_seed
 
 inference_mode_list = ['预训练音色', '3s极速复刻', '跨语种复刻', '自然语言控制']
 instruct_dict = {'预训练音色': '1. 选择预训练音色\n2. 点击生成音频按钮',
@@ -42,13 +43,6 @@ def generate_seed():
     }
 
 
-def set_all_random_seed(seed):
-    random.seed(seed)
-    np.random.seed(seed)
-    torch.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)
-
-
 def postprocess(speech, top_db=60, hop_length=220, win_length=440):
     speech, _ = librosa.effects.trim(
         speech, top_db=top_db,