xinference 0.16.3__py3-none-any.whl → 1.2.1__py3-none-any.whl

xinference/thirdparty/cosyvoice/cli/frontend.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from functools import partial
+import json
 import onnxruntime
 import torch
 import numpy as np
@@ -50,9 +51,13 @@ class CosyVoiceFrontEnd:
         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 torch.cuda.is_available() else "CPUExecutionProvider"])
+        self.speech_tokenizer_session = onnxruntime.InferenceSession(speech_tokenizer_model, sess_options=option,
+                                                                     providers=["CUDAExecutionProvider" if torch.cuda.is_available() else
+                                                                                "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()
@@ -60,10 +65,9 @@ class CosyVoiceFrontEnd:
         if self.use_ttsfrd:
             self.frd = ttsfrd.TtsFrontendEngine()
             ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
-            assert self.frd.initialize('{}/../../pretrained_models/CosyVoice-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)
+            assert self.frd.initialize('{}/../../pretrained_models/CosyVoice-ttsfrd/resource'.format(ROOT_DIR)) is True, \
+                'failed to initialize ttsfrd resource'
+            self.frd.set_lang_type('pinyinvg')
         else:
             self.zh_tn_model = ZhNormalizer(remove_erhua=False, full_to_half=False)
             self.en_tn_model = EnNormalizer()
@@ -75,9 +79,13 @@ class CosyVoiceFrontEnd:
         return text_token, text_token_len
 
     def _extract_speech_token(self, speech):
+        assert speech.shape[1] / 16000 <= 30, 'do not support extract speech token for audio longer than 30s'
         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 = 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
@@ -88,7 +96,8 @@ class CosyVoiceFrontEnd:
                            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 = 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
 
@@ -98,32 +107,34 @@ class CosyVoiceFrontEnd:
         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):
+    def text_normalize(self, text, split=True, text_frontend=True):
+        if text_frontend is False:
+            return [text] if split is True else text
         text = text.strip()
         if contains_chinese(text):
             if self.use_ttsfrd:
-                text = self.frd.get_frd_extra_info(text, 'input')
+                texts = [i["text"] for i in json.loads(self.frd.do_voicegen_frd(text))["sentences"]]
+                text = ''.join(texts)
             else:
                 text = self.zh_tn_model.normalize(text)
-            text = text.replace("\n", "")
-            text = replace_blank(text)
-            text = replace_corner_mark(text)
-            text = text.replace(".", "、")
-            text = text.replace(" - ", "，")
-            text = remove_bracket(text)
-            text = re.sub(r'[，,]+$', '。', text)
-            texts = [i for i in split_paragraph(text, partial(self.tokenizer.encode, allowed_special=self.allowed_special), "zh", token_max_n=80,
-                                                token_min_n=60, merge_len=20,
-                                                comma_split=False)]
+                text = text.replace("\n", "")
+                text = replace_blank(text)
+                text = replace_corner_mark(text)
+                text = text.replace(".", "。")
+                text = text.replace(" - ", "，")
+                text = remove_bracket(text)
+                text = re.sub(r'[，,、]+$', '。', text)
+                texts = list(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:
             if self.use_ttsfrd:
-                text = self.frd.get_frd_extra_info(text, 'input')
+                texts = [i["text"] for i in json.loads(self.frd.do_voicegen_frd(text))["sentences"]]
+                text = ''.join(texts)
             else:
                 text = self.en_tn_model.normalize(text)
-            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)]
+                text = spell_out_number(text, self.inflect_parser)
+                texts = list(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
@@ -134,12 +145,17 @@ class CosyVoiceFrontEnd:
         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):
+    def frontend_zero_shot(self, tts_text, prompt_text, prompt_speech_16k, resample_rate):
         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)
+        prompt_speech_resample = torchaudio.transforms.Resample(orig_freq=16000, new_freq=resample_rate)(prompt_speech_16k)
+        speech_feat, speech_feat_len = self._extract_speech_feat(prompt_speech_resample)
         speech_token, speech_token_len = self._extract_speech_token(prompt_speech_16k)
+        if resample_rate == 24000:
+            # cosyvoice2, force speech_feat % speech_token = 2
+            token_len = min(int(speech_feat.shape[1] / 2), speech_token.shape[1])
+            speech_feat, speech_feat_len[:] = speech_feat[:, :2 * token_len], 2 * token_len
+            speech_token, speech_token_len[:] = speech_token[:, :token_len], token_len
         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,
@@ -149,8 +165,8 @@ class CosyVoiceFrontEnd:
                        '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)
+    def frontend_cross_lingual(self, tts_text, prompt_speech_16k, resample_rate):
+        model_input = self.frontend_zero_shot(tts_text, '', prompt_speech_16k, resample_rate)
         # in cross lingual mode, we remove prompt in llm
         del model_input['prompt_text']
         del model_input['prompt_text_len']
@@ -166,3 +182,34 @@ class CosyVoiceFrontEnd:
         model_input['prompt_text'] = instruct_text_token
         model_input['prompt_text_len'] = instruct_text_token_len
         return model_input
+
+    def frontend_instruct2(self, tts_text, instruct_text, prompt_speech_16k, resample_rate):
+        tts_text_token, tts_text_token_len = self._extract_text_token(tts_text)
+        prompt_text_token, prompt_text_token_len = self._extract_text_token(instruct_text + '<|endofprompt|>')
+        prompt_speech_resample = torchaudio.transforms.Resample(orig_freq=16000, new_freq=resample_rate)(prompt_speech_16k)
+        speech_feat, speech_feat_len = self._extract_speech_feat(prompt_speech_resample)
+        speech_token, speech_token_len = self._extract_speech_token(prompt_speech_16k)
+        if resample_rate == 24000:
+            # cosyvoice2, force speech_feat % speech_token = 2
+            token_len = min(int(speech_feat.shape[1] / 2), speech_token.shape[1])
+            speech_feat, speech_feat_len[:] = speech_feat[:, :2 * token_len], 2 * token_len
+            speech_token, speech_token_len[:] = speech_token[:, :token_len], token_len
+        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,
+                       '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_vc(self, source_speech_16k, prompt_speech_16k, resample_rate):
+        prompt_speech_token, prompt_speech_token_len = self._extract_speech_token(prompt_speech_16k)
+        prompt_speech_resample = torchaudio.transforms.Resample(orig_freq=16000, new_freq=resample_rate)(prompt_speech_16k)
+        prompt_speech_feat, prompt_speech_feat_len = self._extract_speech_feat(prompt_speech_resample)
+        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