add cosyvoice code

cosyvoice/cli/cosyvoice.py

@@ -0,0 +1,83 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+import torch
+from hyperpyyaml import load_hyperpyyaml
+from modelscope import snapshot_download
+from cosyvoice.cli.frontend import CosyVoiceFrontEnd
+from cosyvoice.cli.model import CosyVoiceModel
+
+class CosyVoice:
+
+    def __init__(self, model_dir):
+        instruct = True if '-Instruct' in model_dir else False
+        self.model_dir = model_dir
+        if not os.path.exists(model_dir):
+            model_dir = snapshot_download(model_dir)
+        with open('{}/cosyvoice.yaml'.format(model_dir), 'r') as f:
+            configs = load_hyperpyyaml(f)
+        self.frontend = CosyVoiceFrontEnd(configs['get_tokenizer'],
+                                          configs['feat_extractor'],
+                                          '{}/campplus.onnx'.format(model_dir),
+                                          '{}/speech_tokenizer_v1.onnx'.format(model_dir),
+                                          '{}/spk2info.pt'.format(model_dir),
+                                          instruct,
+                                          configs['allowed_special'])
+        self.model = CosyVoiceModel(configs['llm'], configs['flow'], configs['hift'])
+        self.model.load('{}/llm.pt'.format(model_dir),
+                        '{}/flow.pt'.format(model_dir),
+                        '{}/hift.pt'.format(model_dir))
+        del configs
+
+    def list_avaliable_spks(self):
+        spks = list(self.frontend.spk2info.keys())
+        return spks
+
+    def inference_sft(self, tts_text, spk_id):
+        tts_speeches = []
+        for i in self.frontend.text_normalize(tts_text, split=True):
+            model_input = self.frontend.frontend_sft(i, spk_id)
+            model_output = self.model.inference(**model_input)
+            tts_speeches.append(model_output['tts_speech'])
+        return {'tts_speech': torch.concat(tts_speeches, dim=1)}
+
+    def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k):
+        prompt_text = self.frontend.text_normalize(prompt_text, split=False)
+        tts_speeches = []
+        for i in self.frontend.text_normalize(tts_text, split=True):
+            model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k)
+            model_output = self.model.inference(**model_input)
+            tts_speeches.append(model_output['tts_speech'])
+        return {'tts_speech': torch.concat(tts_speeches, dim=1)}
+
+    def inference_cross_lingual(self, tts_text, prompt_speech_16k):
+        if self.frontend.instruct is True:
+            raise ValueError('{} do not support cross_lingual inference'.format(self.model_dir))
+        tts_speeches = []
+        for i in self.frontend.text_normalize(tts_text, split=True):
+            model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k)
+            model_output = self.model.inference(**model_input)
+            tts_speeches.append(model_output['tts_speech'])
+        return {'tts_speech': torch.concat(tts_speeches, dim=1)}
+
+    def inference_instruct(self, tts_text, spk_id, instruct_text):
+        if self.frontend.instruct is False:
+            raise ValueError('{} do not support instruct inference'.format(self.model_dir))
+        instruct_text = self.frontend.text_normalize(instruct_text, split=False)
+        tts_speeches = []
+        for i in self.frontend.text_normalize(tts_text, split=True):
+            model_input = self.frontend.frontend_instruct(i, spk_id, instruct_text)
+            model_output = self.model.inference(**model_input)
+            tts_speeches.append(model_output['tts_speech'])
+        return {'tts_speech': torch.concat(tts_speeches, dim=1)}

cosyvoice/cli/frontend.py

@@ -0,0 +1,146 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from functools import partial
+import onnxruntime
+import torch
+import numpy as np
+import whisper
+from typing import Callable
+import torchaudio.compliance.kaldi as kaldi
+import torchaudio
+import os
+import inflect
+import ttsfrd
+from cosyvoice.utils.frontend_utils import contains_chinese, replace_blank, replace_corner_mark, remove_bracket, spell_out_number, split_paragraph
+
+
+class CosyVoiceFrontEnd:
+
+    def __init__(self,
+                 get_tokenizer: Callable,
+                 feat_extractor: Callable,
+                 campplus_model: str,
+                 speech_tokenizer_model: str,
+                 spk2info: str = '',
+                 instruct: bool = False,
+                 allowed_special: str = 'all'):
+        self.tokenizer = get_tokenizer()
+        self.feat_extractor = feat_extractor
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        option = onnxruntime.SessionOptions()
+        option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+        option.intra_op_num_threads = 1
+        self.campplus_session = onnxruntime.InferenceSession(campplus_model, sess_options=option, providers=["CPUExecutionProvider"])
+        self.speech_tokenizer_session = onnxruntime.InferenceSession(speech_tokenizer_model, sess_options=option, providers=["CUDAExecutionProvider"])
+        if os.path.exists(spk2info):
+            self.spk2info = torch.load(spk2info, map_location=self.device)
+        self.instruct = instruct
+        self.allowed_special = allowed_special
+        self.inflect_parser = inflect.engine()
+        self.frd = ttsfrd.TtsFrontendEngine()
+        ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+        assert self.frd.initialize('{}/../../pretrained_models/speech_kantts_ttsfrd/resource'.format(ROOT_DIR)) is True, 'failed to initialize ttsfrd resource'
+        self.frd.set_lang_type('pinyin')
+        self.frd.enable_pinyin_mix(True)
+        self.frd.set_breakmodel_index(1)
+
+    def _extract_text_token(self, text):
+        text_token = self.tokenizer.encode(text, allowed_special=self.allowed_special)
+        text_token = torch.tensor([text_token], dtype=torch.int32).to(self.device)
+        text_token_len = torch.tensor([text_token.shape[1]], dtype=torch.int32).to(self.device)
+        return text_token, text_token_len
+
+    def _extract_speech_token(self, speech):
+        feat = whisper.log_mel_spectrogram(speech, n_mels=128)
+        speech_token = self.speech_tokenizer_session.run(None, {self.speech_tokenizer_session.get_inputs()[0].name: feat.detach().cpu().numpy(),
+                                                                self.speech_tokenizer_session.get_inputs()[1].name: np.array([feat.shape[2]], dtype=np.int32)})[0].flatten().tolist()
+        speech_token = torch.tensor([speech_token], dtype=torch.int32).to(self.device)
+        speech_token_len = torch.tensor([speech_token.shape[1]], dtype=torch.int32).to(self.device)
+        return speech_token, speech_token_len
+
+    def _extract_spk_embedding(self, speech):
+        feat = kaldi.fbank(speech,
+                           num_mel_bins=80,
+                           dither=0,
+                           sample_frequency=16000)
+        feat = feat - feat.mean(dim=0, keepdim=True)
+        embedding = self.campplus_session.run(None, {self.campplus_session.get_inputs()[0].name: feat.unsqueeze(dim=0).cpu().numpy()})[0].flatten().tolist()
+        embedding = torch.tensor([embedding]).to(self.device)
+        return embedding
+
+    def _extract_speech_feat(self, speech):
+        speech_feat = self.feat_extractor(speech).squeeze(dim=0).transpose(0, 1).to(self.device)
+        speech_feat = speech_feat.unsqueeze(dim=0)
+        speech_feat_len = torch.tensor([speech_feat.shape[1]], dtype=torch.int32).to(self.device)
+        return speech_feat, speech_feat_len
+
+    def text_normalize(self, text, split=True):
+        text = text.strip()
+        if contains_chinese(text):
+            text = self.frd.get_frd_extra_info(text, 'input').replace("\n", "")
+            text = replace_blank(text)
+            text = replace_corner_mark(text)
+            text = text.replace(".", "、")
+            text = text.replace(" - ", "，")
+            text = remove_bracket(text)
+            texts = [i for i in split_paragraph(text, partial(self.tokenizer.encode, allowed_special=self.allowed_special), "zh", token_max_n=80,
+                                                token_min_n=60, merge_len=20,
+                                                comma_split=False)]
+        else:
+            text = spell_out_number(text, self.inflect_parser)
+            texts = [i for i in split_paragraph(text, partial(self.tokenizer.encode, allowed_special=self.allowed_special), "en", token_max_n=80,
+                                                token_min_n=60, merge_len=20,
+                                                comma_split=False)]
+        if split is False:
+            return text
+        return texts
+
+    def frontend_sft(self, tts_text, spk_id):
+        tts_text_token, tts_text_token_len = self._extract_text_token(tts_text)
+        embedding = self.spk2info[spk_id]['embedding']
+        model_input = {'text': tts_text_token, 'text_len': tts_text_token_len, 'llm_embedding': embedding, 'flow_embedding': embedding}
+        return model_input
+
+    def frontend_zero_shot(self, tts_text, prompt_text, prompt_speech_16k):
+        tts_text_token, tts_text_token_len = self._extract_text_token(tts_text)
+        prompt_text_token, prompt_text_token_len = self._extract_text_token(prompt_text)
+        prompt_speech_22050 = torchaudio.transforms.Resample(orig_freq=16000, new_freq=22050)(prompt_speech_16k)
+        speech_feat, speech_feat_len = self._extract_speech_feat(prompt_speech_22050)
+        speech_token, speech_token_len = self._extract_speech_token(prompt_speech_16k)
+        embedding = self._extract_spk_embedding(prompt_speech_16k)
+        model_input = {'text': tts_text_token, 'text_len': tts_text_token_len,
+                       'prompt_text': prompt_text_token, 'prompt_text_len': prompt_text_token_len,
+                       'llm_prompt_speech_token': speech_token, 'llm_prompt_speech_token_len': speech_token_len,
+                       'flow_prompt_speech_token': speech_token, 'flow_prompt_speech_token_len': speech_token_len,
+                       'prompt_speech_feat': speech_feat, 'prompt_speech_feat_len': speech_feat_len,
+                       'llm_embedding': embedding, 'flow_embedding': embedding}
+        return model_input
+
+    def frontend_cross_lingual(self, tts_text, prompt_speech_16k):
+        model_input = self.frontend_zero_shot(tts_text, '', prompt_speech_16k)
+        # in cross lingual mode, we remove prompt in llm
+        del model_input['prompt_text']
+        del model_input['prompt_text_len']
+        del model_input['llm_prompt_speech_token']
+        del model_input['llm_prompt_speech_token_len']
+        return model_input
+
+    def frontend_instruct(self, tts_text, spk_id, instruct_text):
+        model_input = self.frontend_sft(tts_text, spk_id)
+        # in instruct mode, we remove spk_embedding in llm due to information leakage
+        del model_input['llm_embedding']
+        instruct_text_token, instruct_text_token_len = self._extract_text_token(instruct_text + '<endofprompt>')
+        model_input['prompt_text'] = instruct_text_token
+        model_input['prompt_text_len'] = instruct_text_token_len
+        return model_input

cosyvoice/cli/model.py

@@ -0,0 +1,59 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+class CosyVoiceModel:
+
+    def __init__(self,
+                 llm: torch.nn.Module,
+                 flow: torch.nn.Module,
+                 hift: torch.nn.Module):
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self.llm = llm
+        self.flow = flow
+        self.hift = hift
+
+    def load(self, llm_model, flow_model, hift_model):
+        self.llm.load_state_dict(torch.load(llm_model, map_location=self.device))
+        self.llm.to(self.device).eval()
+        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device))
+        self.flow.to(self.device).eval()
+        self.hift.load_state_dict(torch.load(hift_model, map_location=self.device))
+        self.hift.to(self.device).eval()
+
+    def inference(self, text, text_len, flow_embedding, llm_embedding=torch.zeros(0, 192),
+                  prompt_text=torch.zeros(1, 0, dtype=torch.int32), prompt_text_len=torch.zeros(1, dtype=torch.int32),
+                  llm_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), llm_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
+                  flow_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), flow_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
+                  prompt_speech_feat=torch.zeros(1, 0, 80), prompt_speech_feat_len=torch.zeros(1, dtype=torch.int32)):
+        tts_speech_token = self.llm.inference(text=text.to(self.device),
+                                              text_len=text_len.to(self.device),
+                                              prompt_text=prompt_text.to(self.device),
+                                              prompt_text_len=prompt_text_len.to(self.device),
+                                              prompt_speech_token=llm_prompt_speech_token.to(self.device),
+                                              prompt_speech_token_len=llm_prompt_speech_token_len.to(self.device),
+                                              embedding=llm_embedding.to(self.device),
+                                              beam_size=1,
+                                              sampling=25,
+                                              max_token_text_ratio=30,
+                                              min_token_text_ratio=3)
+        tts_mel = self.flow.inference(token=tts_speech_token,
+                                      token_len=torch.tensor([tts_speech_token.size(1)], dtype=torch.int32).to(self.device),
+                                      prompt_token=flow_prompt_speech_token.to(self.device),
+                                      prompt_token_len=flow_prompt_speech_token_len.to(self.device),
+                                      prompt_feat=prompt_speech_feat.to(self.device),
+                                      prompt_feat_len=prompt_speech_feat_len.to(self.device),
+                                      embedding=flow_embedding.to(self.device))
+        tts_speech = self.hift.inference(mel=tts_mel).cpu()
+        return {'tts_speech': tts_speech}