xinference 1.0.1__py3-none-any.whl → 1.2.1__py3-none-any.whl

xinference/thirdparty/cosyvoice/bin/inference.py

@@ -18,16 +18,15 @@ import argparse
 import logging
 logging.getLogger('matplotlib').setLevel(logging.WARNING)
 import os
-
 import torch
 from torch.utils.data import DataLoader
 import torchaudio
 from hyperpyyaml import load_hyperpyyaml
 from tqdm import tqdm
 from cosyvoice.cli.model import CosyVoiceModel
-
 from cosyvoice.dataset.dataset import Dataset
 
+
 def get_args():
     parser = argparse.ArgumentParser(description='inference with your model')
     parser.add_argument('--config', required=True, help='config file')
@@ -66,7 +65,8 @@ def main():
     model = CosyVoiceModel(configs['llm'], configs['flow'], configs['hift'])
     model.load(args.llm_model, args.flow_model, args.hifigan_model)
 
-    test_dataset = Dataset(args.prompt_data, data_pipeline=configs['data_pipeline'], mode='inference', shuffle=False, partition=False, tts_file=args.tts_text, prompt_utt2data=args.prompt_utt2data)
+    test_dataset = Dataset(args.prompt_data, data_pipeline=configs['data_pipeline'], mode='inference', shuffle=False, partition=False,
+                           tts_file=args.tts_text, prompt_utt2data=args.prompt_utt2data)
     test_data_loader = DataLoader(test_dataset, batch_size=None, num_workers=0)
 
     del configs
@@ -74,13 +74,11 @@ def main():
     fn = os.path.join(args.result_dir, 'wav.scp')
     f = open(fn, 'w')
     with torch.no_grad():
-        for batch_idx, batch in tqdm(enumerate(test_data_loader)):
+        for _, batch in tqdm(enumerate(test_data_loader)):
             utts = batch["utts"]
             assert len(utts) == 1, "inference mode only support batchsize 1"
-            text = batch["text"]
             text_token = batch["text_token"].to(device)
             text_token_len = batch["text_token_len"].to(device)
-            tts_text = batch["tts_text"]
             tts_index = batch["tts_index"]
             tts_text_token = batch["tts_text_token"].to(device)
             tts_text_token_len = batch["tts_text_token_len"].to(device)
@@ -101,7 +99,7 @@ def main():
                                'prompt_speech_feat': speech_feat, 'prompt_speech_feat_len': speech_feat_len,
                                'llm_embedding': utt_embedding, 'flow_embedding': utt_embedding}
             tts_speeches = []
-            for model_output in model.inference(**model_input):
+            for model_output in model.tts(**model_input):
                 tts_speeches.append(model_output['tts_speech'])
             tts_speeches = torch.concat(tts_speeches, dim=1)
             tts_key = '{}_{}'.format(utts[0], tts_index[0])

xinference/thirdparty/cosyvoice/bin/train.py

@@ -18,6 +18,7 @@ import datetime
 import logging
 logging.getLogger('matplotlib').setLevel(logging.WARNING)
 from copy import deepcopy
+import os
 import torch
 import torch.distributed as dist
 import deepspeed
@@ -67,13 +68,17 @@ def get_args():
                         action='store_true',
                         default=False,
                         help='Use pinned memory buffers used for reading')
+    parser.add_argument('--use_amp',
+                        action='store_true',
+                        default=False,
+                        help='Use automatic mixed precision training')
     parser.add_argument('--deepspeed.save_states',
                         dest='save_states',
                         default='model_only',
                         choices=['model_only', 'model+optimizer'],
                         help='save model/optimizer states')
     parser.add_argument('--timeout',
-                        default=30,
+                        default=60,
                         type=int,
                         help='timeout (in seconds) of cosyvoice_join.')
     parser = deepspeed.add_config_arguments(parser)
@@ -86,10 +91,16 @@ def main():
     args = get_args()
     logging.basicConfig(level=logging.DEBUG,
                         format='%(asctime)s %(levelname)s %(message)s')
+    # gan train has some special initialization logic
+    gan = True if args.model == 'hifigan' else False
 
-    override_dict = {k: None for k in ['llm', 'flow', 'hift'] if k != args.model}
+    override_dict = {k: None for k in ['llm', 'flow', 'hift', 'hifigan'] if k != args.model}
+    if gan is True:
+        override_dict.pop('hift')
     with open(args.config, 'r') as f:
         configs = load_hyperpyyaml(f, overrides=override_dict)
+    if gan is True:
+        configs['train_conf'] = configs['train_conf_gan']
     configs['train_conf'].update(vars(args))
 
     # Init env for ddp
@@ -97,7 +108,7 @@ def main():
 
     # Get dataset & dataloader
     train_dataset, cv_dataset, train_data_loader, cv_data_loader = \
-        init_dataset_and_dataloader(args, configs)
+        init_dataset_and_dataloader(args, configs, gan)
 
     # Do some sanity checks and save config to arsg.model_dir
     configs = check_modify_and_save_config(args, configs)
@@ -107,30 +118,53 @@ def main():
 
     # load checkpoint
     model = configs[args.model]
+    start_step, start_epoch = 0, -1
     if args.checkpoint is not None:
-        model.load_state_dict(torch.load(args.checkpoint, map_location='cpu'))
+        if os.path.exists(args.checkpoint):
+            state_dict = torch.load(args.checkpoint, map_location='cpu')
+            model.load_state_dict(state_dict, strict=False)
+            if 'step' in state_dict:
+                start_step = state_dict['step']
+            if 'epoch' in state_dict:
+                start_epoch = state_dict['epoch']
+        else:
+            logging.warning('checkpoint {} do not exsist!'.format(args.checkpoint))
 
     # Dispatch model from cpu to gpu
     model = wrap_cuda_model(args, model)
 
     # Get optimizer & scheduler
-    model, optimizer, scheduler = init_optimizer_and_scheduler(args, configs, model)
+    model, optimizer, scheduler, optimizer_d, scheduler_d = init_optimizer_and_scheduler(args, configs, model, gan)
+    scheduler.set_step(start_step)
+    if scheduler_d is not None:
+        scheduler_d.set_step(start_step)
 
     # Save init checkpoints
     info_dict = deepcopy(configs['train_conf'])
+    info_dict['step'] = start_step
+    info_dict['epoch'] = start_epoch
     save_model(model, 'init', info_dict)
 
     # Get executor
-    executor = Executor()
+    executor = Executor(gan=gan)
+    executor.step = start_step
 
+    # Init scaler, used for pytorch amp mixed precision training
+    scaler = torch.cuda.amp.GradScaler() if args.use_amp else None
+    print('start step {} start epoch {}'.format(start_step, start_epoch))
     # Start training loop
-    for epoch in range(info_dict['max_epoch']):
+    for epoch in range(start_epoch + 1, info_dict['max_epoch']):
         executor.epoch = epoch
         train_dataset.set_epoch(epoch)
         dist.barrier()
         group_join = dist.new_group(backend="gloo", timeout=datetime.timedelta(seconds=args.timeout))
-        executor.train_one_epoc(model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, group_join)
+        if gan is True:
+            executor.train_one_epoc_gan(model, optimizer, scheduler, optimizer_d, scheduler_d, train_data_loader, cv_data_loader,
+                                        writer, info_dict, scaler, group_join)
+        else:
+            executor.train_one_epoc(model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, scaler, group_join)
         dist.destroy_process_group(group_join)
 
+
 if __name__ == '__main__':
     main()

xinference/thirdparty/cosyvoice/cli/cosyvoice.py

@@ -13,15 +13,18 @@
 # limitations under the License.
 import os
 import time
+from tqdm import tqdm
 from hyperpyyaml import load_hyperpyyaml
 from modelscope import snapshot_download
+import torch
 from cosyvoice.cli.frontend import CosyVoiceFrontEnd
-from cosyvoice.cli.model import CosyVoiceModel
+from cosyvoice.cli.model import CosyVoiceModel, CosyVoice2Model
 from cosyvoice.utils.file_utils import logging
 
+
 class CosyVoice:
 
-    def __init__(self, model_dir, load_jit=True):
+    def __init__(self, model_dir, load_jit=True, load_onnx=False, fp16=True):
         instruct = True if '-Instruct' in model_dir else False
         self.model_dir = model_dir
         if not os.path.exists(model_dir):
@@ -35,65 +38,133 @@ class CosyVoice:
                                           '{}/spk2info.pt'.format(model_dir),
                                           instruct,
                                           configs['allowed_special'])
-        self.model = CosyVoiceModel(configs['llm'], configs['flow'], configs['hift'])
+        self.sample_rate = configs['sample_rate']
+        if torch.cuda.is_available() is False and (fp16 is True or load_jit is True):
+            load_jit = False
+            fp16 = False
+            logging.warning('cpu do not support fp16 and jit, force set to False')
+        self.model = CosyVoiceModel(configs['llm'], configs['flow'], configs['hift'], fp16)
         self.model.load('{}/llm.pt'.format(model_dir),
                         '{}/flow.pt'.format(model_dir),
                         '{}/hift.pt'.format(model_dir))
         if load_jit:
             self.model.load_jit('{}/llm.text_encoder.fp16.zip'.format(model_dir),
-                                    '{}/llm.llm.fp16.zip'.format(model_dir))
+                                '{}/llm.llm.fp16.zip'.format(model_dir),
+                                '{}/flow.encoder.fp32.zip'.format(model_dir))
+        if load_onnx:
+            self.model.load_onnx('{}/flow.decoder.estimator.fp32.onnx'.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, stream=False):
-        for i in self.frontend.text_normalize(tts_text, split=True):
+    def inference_sft(self, tts_text, spk_id, stream=False, speed=1.0, text_frontend=True):
+        for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
             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):
-                speech_len = model_output['tts_speech'].shape[1] / 22050
+            for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
+                speech_len = model_output['tts_speech'].shape[1] / self.sample_rate
                 logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
                 yield model_output
                 start_time = time.time()
 
-    def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k, stream=False):
-        prompt_text = self.frontend.text_normalize(prompt_text, split=False)
-        for i in self.frontend.text_normalize(tts_text, split=True):
-            model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k)
+    def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True):
+        prompt_text = self.frontend.text_normalize(prompt_text, split=False, text_frontend=text_frontend)
+        for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
+            if len(i) < 0.5 * len(prompt_text):
+                logging.warning('synthesis text {} too short than prompt text {}, this may lead to bad performance'.format(i, prompt_text))
+            model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k, self.sample_rate)
             start_time = time.time()
             logging.info('synthesis text {}'.format(i))
-            for model_output in self.model.inference(**model_input, stream=stream):
-                speech_len = model_output['tts_speech'].shape[1] / 22050
+            for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
+                speech_len = model_output['tts_speech'].shape[1] / self.sample_rate
                 logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
                 yield model_output
                 start_time = time.time()
 
-    def inference_cross_lingual(self, tts_text, prompt_speech_16k, stream=False):
-        if self.frontend.instruct is True:
+    def inference_cross_lingual(self, tts_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True):
+        if self.frontend.instruct is True and isinstance(self.model, CosyVoiceModel):
             raise ValueError('{} do not support cross_lingual inference'.format(self.model_dir))
-        for i in self.frontend.text_normalize(tts_text, split=True):
-            model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k)
+        for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
+            model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k, self.sample_rate)
             start_time = time.time()
             logging.info('synthesis text {}'.format(i))
-            for model_output in self.model.inference(**model_input, stream=stream):
-                speech_len = model_output['tts_speech'].shape[1] / 22050
+            for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
+                speech_len = model_output['tts_speech'].shape[1] / self.sample_rate
                 logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
                 yield model_output
                 start_time = time.time()
 
-    def inference_instruct(self, tts_text, spk_id, instruct_text, stream=False):
+    def inference_instruct(self, tts_text, spk_id, instruct_text, stream=False, speed=1.0, text_frontend=True):
+        assert isinstance(self.model, CosyVoiceModel)
         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)
-        for i in self.frontend.text_normalize(tts_text, split=True):
+        instruct_text = self.frontend.text_normalize(instruct_text, split=False, text_frontend=text_frontend)
+        for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
             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):
-                speech_len = model_output['tts_speech'].shape[1] / 22050
+            for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
+                speech_len = model_output['tts_speech'].shape[1] / self.sample_rate
+                logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
+                yield model_output
+                start_time = time.time()
+
+    def inference_instruct2(self, tts_text, instruct_text, prompt_speech_16k, stream=False, speed=1.0, text_frontend=True):
+        assert isinstance(self.model, CosyVoice2Model)
+        for i in tqdm(self.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
+            model_input = self.frontend.frontend_instruct2(i, instruct_text, prompt_speech_16k, self.sample_rate)
+            start_time = time.time()
+            logging.info('synthesis text {}'.format(i))
+            for model_output in self.model.tts(**model_input, stream=stream, speed=speed):
+                speech_len = model_output['tts_speech'].shape[1] / self.sample_rate
                 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, self.sample_rate)
+        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] / self.sample_rate
+            logging.info('yield speech len {}, rtf {}'.format(speech_len, (time.time() - start_time) / speech_len))
+            yield model_output
+            start_time = time.time()
+
+
+class CosyVoice2(CosyVoice):
+
+    def __init__(self, model_dir, load_jit=False, load_onnx=False, load_trt=False):
+        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, overrides={'qwen_pretrain_path': os.path.join(model_dir, 'CosyVoice-BlankEN')})
+        self.frontend = CosyVoiceFrontEnd(configs['get_tokenizer'],
+                                          configs['feat_extractor'],
+                                          '{}/campplus.onnx'.format(model_dir),
+                                          '{}/speech_tokenizer_v2.onnx'.format(model_dir),
+                                          '{}/spk2info.pt'.format(model_dir),
+                                          instruct,
+                                          configs['allowed_special'])
+        self.sample_rate = configs['sample_rate']
+        if torch.cuda.is_available() is False and load_jit is True:
+            load_jit = False
+            logging.warning('cpu do not support jit, force set to False')
+        self.model = CosyVoice2Model(configs['llm'], configs['flow'], configs['hift'])
+        self.model.load('{}/llm.pt'.format(model_dir),
+                        '{}/flow.pt'.format(model_dir),
+                        '{}/hift.pt'.format(model_dir))
+        if load_jit:
+            self.model.load_jit('{}/flow.encoder.fp32.zip'.format(model_dir))
+        if load_trt is True and load_onnx is True:
+            load_onnx = False
+            logging.warning('can not set both load_trt and load_onnx to True, force set load_onnx to False')
+        if load_onnx:
+            self.model.load_onnx('{}/flow.decoder.estimator.fp32.onnx'.format(model_dir))
+        if load_trt:
+            self.model.load_trt('{}/flow.decoder.estimator.fp16.Volta.plan'.format(model_dir))
+        del configs

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