xinference 1.0.1__py3-none-any.whl → 1.1.1__py3-none-any.whl

xinference/thirdparty/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
@@ -102,6 +104,7 @@ def init_weights(m, mean=0.0, std=0.01):
     if classname.find("Conv") != -1:
         m.weight.data.normal_(mean, std)
 
+
 # Repetition Aware Sampling in VALL-E 2
 def ras_sampling(weighted_scores, decoded_tokens, sampling, top_p=0.8, top_k=25, win_size=10, tau_r=0.1):
     top_ids = nucleus_sampling(weighted_scores, top_p=top_p, top_k=top_k)
@@ -110,6 +113,7 @@ def ras_sampling(weighted_scores, decoded_tokens, sampling, top_p=0.8, top_k=25,
         top_ids = random_sampling(weighted_scores, decoded_tokens, sampling)
     return top_ids
 
+
 def nucleus_sampling(weighted_scores, top_p=0.8, top_k=25):
     prob, indices = [], []
     cum_prob = 0.0
@@ -127,13 +131,36 @@ def nucleus_sampling(weighted_scores, top_p=0.8, top_k=25):
     top_ids = indices[prob.multinomial(1, replacement=True)]
     return top_ids
 
+
 def random_sampling(weighted_scores, decoded_tokens, sampling):
     top_ids = weighted_scores.softmax(dim=0).multinomial(1, replacement=True)
     return top_ids
 
+
 def fade_in_out(fade_in_mel, fade_out_mel, window):
     device = fade_in_mel.device
     fade_in_mel, fade_out_mel = fade_in_mel.cpu(), fade_out_mel.cpu()
     mel_overlap_len = int(window.shape[0] / 2)
-    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:]
+    if fade_in_mel.device == torch.device('cpu'):
+        fade_in_mel = fade_in_mel.clone()
+    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)
+
+
+def mask_to_bias(mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
+    assert mask.dtype == torch.bool
+    assert dtype in [torch.float32, torch.bfloat16, torch.float16]
+    mask = mask.to(dtype)
+    # attention mask bias
+    # NOTE(Mddct): torch.finfo jit issues
+    #     chunk_masks = (1.0 - chunk_masks) * torch.finfo(dtype).min
+    mask = (1.0 - mask) * torch.finfo(dtype).min
+    return mask

xinference/thirdparty/cosyvoice/utils/executor.py

@@ -25,13 +25,14 @@ from cosyvoice.utils.train_utils import update_parameter_and_lr, log_per_step, l
 
 class Executor:
 
-    def __init__(self):
+    def __init__(self, gan: bool = False):
+        self.gan = gan
         self.step = 0
         self.epoch = 0
         self.rank = int(os.environ.get('RANK', 0))
         self.device = torch.device('cuda:{}'.format(self.rank))
 
-    def train_one_epoc(self, model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, group_join):
+    def train_one_epoc(self, model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, scaler, group_join):
         ''' Train one epoch
         '''
 
@@ -64,13 +65,72 @@ class Executor:
                     context = nullcontext
 
                 with context():
-                    info_dict = batch_forward(model, batch_dict, info_dict)
-                    info_dict = batch_backward(model, info_dict)
+                    info_dict = batch_forward(model, batch_dict, scaler, info_dict)
+                    info_dict = batch_backward(model, scaler, info_dict)
 
-                info_dict = update_parameter_and_lr(model, optimizer, scheduler, info_dict)
+                info_dict = update_parameter_and_lr(model, optimizer, scheduler, scaler, info_dict)
                 log_per_step(writer, info_dict)
                 # NOTE specify save_per_step in cosyvoice.yaml if you want to enable step save
-                if info_dict['save_per_step'] > 0 and (self.step + 1) % info_dict['save_per_step'] == 0 and (batch_idx + 1) % info_dict["accum_grad"] == 0:
+                if info_dict['save_per_step'] > 0 and (self.step + 1) % info_dict['save_per_step'] == 0 and \
+                   (batch_idx + 1) % info_dict["accum_grad"] == 0:
+                    dist.barrier()
+                    self.cv(model, cv_data_loader, writer, info_dict, on_batch_end=False)
+                    model.train()
+                if (batch_idx + 1) % info_dict["accum_grad"] == 0:
+                    self.step += 1
+        dist.barrier()
+        self.cv(model, cv_data_loader, writer, info_dict, on_batch_end=True)
+
+    def train_one_epoc_gan(self, model, optimizer, scheduler, optimizer_d, scheduler_d, train_data_loader, cv_data_loader,
+                           writer, info_dict, scaler, group_join):
+        ''' Train one epoch
+        '''
+
+        lr = optimizer.param_groups[0]['lr']
+        logging.info('Epoch {} TRAIN info lr {} rank {}'.format(self.epoch, lr, self.rank))
+        logging.info('using accumulate grad, new batch size is {} times'
+                     ' larger than before'.format(info_dict['accum_grad']))
+        # A context manager to be used in conjunction with an instance of
+        # torch.nn.parallel.DistributedDataParallel to be able to train
+        # with uneven inputs across participating processes.
+        model.train()
+        model_context = model.join if info_dict['train_engine'] == 'torch_ddp' else nullcontext
+        with model_context():
+            for batch_idx, batch_dict in enumerate(train_data_loader):
+                info_dict["tag"] = "TRAIN"
+                info_dict["step"] = self.step
+                info_dict["epoch"] = self.epoch
+                info_dict["batch_idx"] = batch_idx
+                if cosyvoice_join(group_join, info_dict):
+                    break
+
+                # Disable gradient synchronizations across DDP processes.
+                # Within this context, gradients will be accumulated on module
+                # variables, which will later be synchronized.
+                if info_dict['train_engine'] == 'torch_ddp' and (batch_idx + 1) % info_dict["accum_grad"] != 0:
+                    context = model.no_sync
+                # Used for single gpu training and DDP gradient synchronization
+                # processes.
+                else:
+                    context = nullcontext
+
+                with context():
+                    batch_dict['turn'] = 'discriminator'
+                    info_dict = batch_forward(model, batch_dict, scaler, info_dict)
+                    info_dict = batch_backward(model, scaler, info_dict)
+                info_dict = update_parameter_and_lr(model, optimizer_d, scheduler_d, scaler, info_dict)
+                optimizer.zero_grad()
+                log_per_step(writer, info_dict)
+                with context():
+                    batch_dict['turn'] = 'generator'
+                    info_dict = batch_forward(model, batch_dict, scaler, info_dict)
+                    info_dict = batch_backward(model, scaler, info_dict)
+                info_dict = update_parameter_and_lr(model, optimizer, scheduler, scaler, info_dict)
+                optimizer_d.zero_grad()
+                log_per_step(writer, info_dict)
+                # NOTE specify save_per_step in cosyvoice.yaml if you want to enable step save
+                if info_dict['save_per_step'] > 0 and (self.step + 1) % info_dict['save_per_step'] == 0 and \
+                   (batch_idx + 1) % info_dict["accum_grad"] == 0:
                     dist.barrier()
                     self.cv(model, cv_data_loader, writer, info_dict, on_batch_end=False)
                     model.train()
@@ -95,7 +155,9 @@ class Executor:
             num_utts = len(batch_dict["utts"])
             total_num_utts += num_utts
 
-            info_dict = batch_forward(model, batch_dict, info_dict)
+            if self.gan is True:
+                batch_dict['turn'] = 'generator'
+            info_dict = batch_forward(model, batch_dict, None, info_dict)
 
             for k, v in info_dict['loss_dict'].items():
                 if k not in total_loss_dict:

xinference/thirdparty/cosyvoice/utils/file_utils.py

@@ -28,6 +28,7 @@ def read_lists(list_file):
             lists.append(line.strip())
     return lists
 
+
 def read_json_lists(list_file):
     lists = read_lists(list_file)
     results = {}
@@ -36,6 +37,7 @@ def read_json_lists(list_file):
             results.update(json.load(fin))
     return results
 
+
 def load_wav(wav, target_sr):
     speech, sample_rate = torchaudio.load(wav)
     speech = speech.mean(dim=0, keepdim=True)
@@ -43,15 +45,3 @@ def load_wav(wav, target_sr):
         assert sample_rate > target_sr, 'wav sample rate {} must be greater than {}'.format(sample_rate, target_sr)
         speech = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_sr)(speech)
     return speech
-
-def speed_change(waveform, sample_rate, speed_factor: str):
-    effects = [
-        ["tempo", speed_factor],  # speed_factor
-        ["rate", f"{sample_rate}"]
-    ]
-    augmented_waveform, new_sample_rate = torchaudio.sox_effects.apply_effects_tensor(
-        waveform,
-        sample_rate,
-        effects
-    )
-    return augmented_waveform, new_sample_rate

xinference/thirdparty/cosyvoice/utils/frontend_utils.py

@@ -15,6 +15,7 @@
 import re
 chinese_char_pattern = re.compile(r'[\u4e00-\u9fff]+')
 
+
 # whether contain chinese character
 def contains_chinese(text):
     return bool(chinese_char_pattern.search(text))
@@ -79,6 +80,13 @@ def split_paragraph(text: str, tokenize, lang="zh", token_max_n=80, token_min_n=
         pounc = ['.', '?', '!', ';', ':']
     if comma_split:
         pounc.extend(['，', ','])
+
+    if text[-1] not in pounc:
+        if lang == "zh":
+            text += "。"
+        else:
+            text += "."
+
     st = 0
     utts = []
     for i, c in enumerate(text):
@@ -91,11 +99,7 @@ def split_paragraph(text: str, tokenize, lang="zh", token_max_n=80, token_min_n=
                 st = i + 2
             else:
                 st = i + 1
-    if len(utts) == 0:
-        if lang == "zh":
-            utts.append(text + '。')
-        else:
-            utts.append(text + '.')
+
     final_utts = []
     cur_utt = ""
     for utt in utts:

xinference/thirdparty/cosyvoice/utils/losses.py

@@ -0,0 +1,20 @@
+import torch
+import torch.nn.functional as F
+
+
+def tpr_loss(disc_real_outputs, disc_generated_outputs, tau):
+    loss = 0
+    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+        m_DG = torch.median((dr - dg))
+        L_rel = torch.mean((((dr - dg) - m_DG) ** 2)[dr < dg + m_DG])
+        loss += tau - F.relu(tau - L_rel)
+    return loss
+
+
+def mel_loss(real_speech, generated_speech, mel_transforms):
+    loss = 0
+    for transform in mel_transforms:
+        mel_r = transform(real_speech)
+        mel_g = transform(generated_speech)
+        loss += F.l1_loss(mel_g, mel_r)
+    return loss

xinference/thirdparty/cosyvoice/utils/scheduler.py

@@ -567,8 +567,7 @@ class NoamAnnealing(_LRScheduler):
                  min_lr=0.0,
                  last_epoch=-1):
         self._normalize = d_model**(-0.5)
-        assert not (warmup_steps is not None
-                    and warmup_ratio is not None), \
+        assert not (warmup_steps is not None and warmup_ratio is not None), \
             "Either use particular number of step or ratio"
         assert warmup_ratio is None or max_steps is not None, \
             "If there is a ratio, there should be a total steps"

xinference/thirdparty/cosyvoice/utils/train_utils.py

@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from contextlib import nullcontext
 import logging
 import os
 import torch
@@ -51,9 +50,10 @@ def init_distributed(args):
     return world_size, local_rank, rank
 
 
-def init_dataset_and_dataloader(args, configs):
-    train_dataset = Dataset(args.train_data, data_pipeline=configs['data_pipeline'], mode='train', shuffle=True, partition=True)
-    cv_dataset = Dataset(args.cv_data, data_pipeline=configs['data_pipeline'], mode='train', shuffle=False, partition=False)
+def init_dataset_and_dataloader(args, configs, gan):
+    data_pipeline = configs['data_pipeline_gan'] if gan is True else configs['data_pipeline']
+    train_dataset = Dataset(args.train_data, data_pipeline=data_pipeline, mode='train', gan=gan, shuffle=True, partition=True)
+    cv_dataset = Dataset(args.cv_data, data_pipeline=data_pipeline, mode='train', gan=gan, shuffle=False, partition=False)
 
     # do not use persistent_workers=True, as whisper tokenizer opens tiktoken file each time when the for loop starts
     train_data_loader = DataLoader(train_dataset,
@@ -69,7 +69,6 @@ def init_dataset_and_dataloader(args, configs):
     return train_dataset, cv_dataset, train_data_loader, cv_data_loader
 
 
-
 def check_modify_and_save_config(args, configs):
     if args.train_engine == "torch_ddp":
         configs['train_conf']["dtype"] = 'fp32'
@@ -84,7 +83,8 @@ def check_modify_and_save_config(args, configs):
             configs['train_conf']["dtype"] = "fp32"
         assert ds_configs["train_micro_batch_size_per_gpu"] == 1
         # if use deepspeed, override ddp config
-        configs['train_conf']['save_per_step'] = int(configs['train_conf']['save_per_step'] * configs['train_conf']['accum_grad'] / ds_configs["gradient_accumulation_steps"])
+        configs['train_conf']['save_per_step'] = int(configs['train_conf']['save_per_step'] *
+                                                     configs['train_conf']['accum_grad'] / ds_configs["gradient_accumulation_steps"])
         configs['train_conf']['accum_grad'] = ds_configs["gradient_accumulation_steps"]
         configs['train_conf']['grad_clip'] = ds_configs["gradient_clipping"]
         configs['train_conf']['log_interval'] = ds_configs["steps_per_print"]
@@ -108,38 +108,80 @@ def wrap_cuda_model(args, model):
     return model
 
 
-def init_optimizer_and_scheduler(args, configs, model):
-    if configs['train_conf']['optim'] == 'adam':
-        optimizer = optim.Adam(model.parameters(), **configs['train_conf']['optim_conf'])
-    elif configs['train_conf']['optim'] == 'adamw':
-        optimizer = optim.AdamW(model.parameters(), **configs['train_conf']['optim_conf'])
-    else:
-        raise ValueError("unknown optimizer: " + configs['train_conf'])
-
-    if configs['train_conf']['scheduler'] == 'warmuplr':
-        scheduler_type = WarmupLR
-        scheduler = WarmupLR(optimizer, **configs['train_conf']['scheduler_conf'])
-    elif configs['train_conf']['scheduler'] == 'NoamHoldAnnealing':
-        scheduler_type = NoamHoldAnnealing
-        scheduler = NoamHoldAnnealing(optimizer, **configs['train_conf']['scheduler_conf'])
-    elif configs['train_conf']['scheduler'] == 'constantlr':
-        scheduler_type = ConstantLR
-        scheduler = ConstantLR(optimizer)
-    else:
-        raise ValueError("unknown scheduler: " + configs['train_conf'])
+def init_optimizer_and_scheduler(args, configs, model, gan):
+    if gan is False:
+        if configs['train_conf']['optim'] == 'adam':
+            optimizer = optim.Adam(model.parameters(), **configs['train_conf']['optim_conf'])
+        elif configs['train_conf']['optim'] == 'adamw':
+            optimizer = optim.AdamW(model.parameters(), **configs['train_conf']['optim_conf'])
+        else:
+            raise ValueError("unknown optimizer: " + configs['train_conf'])
+
+        if configs['train_conf']['scheduler'] == 'warmuplr':
+            scheduler_type = WarmupLR
+            scheduler = WarmupLR(optimizer, **configs['train_conf']['scheduler_conf'])
+        elif configs['train_conf']['scheduler'] == 'NoamHoldAnnealing':
+            scheduler_type = NoamHoldAnnealing
+            scheduler = NoamHoldAnnealing(optimizer, **configs['train_conf']['scheduler_conf'])
+        elif configs['train_conf']['scheduler'] == 'constantlr':
+            scheduler_type = ConstantLR
+            scheduler = ConstantLR(optimizer)
+        else:
+            raise ValueError("unknown scheduler: " + configs['train_conf'])
+
+        # use deepspeed optimizer for speedup
+        if args.train_engine == "deepspeed":
+            def scheduler(opt):
+                return scheduler_type(opt, **configs['train_conf']['scheduler_conf'])
+            model, optimizer, _, scheduler = deepspeed.initialize(
+                args=args,
+                model=model,
+                optimizer=None,
+                lr_scheduler=scheduler,
+                model_parameters=model.parameters())
+
+        optimizer_d, scheduler_d = None, None
 
-    # use deepspeed optimizer for speedup
-    if args.train_engine == "deepspeed":
-        def scheduler(opt):
-            return scheduler_type(opt, **configs['train_conf']['scheduler_conf'])
-        model, optimizer, _, scheduler = deepspeed.initialize(
-            args=args,
-            model=model,
-            optimizer=None,
-            lr_scheduler=scheduler,
-            model_parameters=model.parameters())
+    else:
+        # currently we wrap generator and discriminator in one model, so we cannot use deepspeed
+        if configs['train_conf']['optim'] == 'adam':
+            optimizer = optim.Adam(model.module.generator.parameters(), **configs['train_conf']['optim_conf'])
+        elif configs['train_conf']['optim'] == 'adamw':
+            optimizer = optim.AdamW(model.module.generator.parameters(), **configs['train_conf']['optim_conf'])
+        else:
+            raise ValueError("unknown optimizer: " + configs['train_conf'])
+
+        if configs['train_conf']['scheduler'] == 'warmuplr':
+            scheduler_type = WarmupLR
+            scheduler = WarmupLR(optimizer, **configs['train_conf']['scheduler_conf'])
+        elif configs['train_conf']['scheduler'] == 'NoamHoldAnnealing':
+            scheduler_type = NoamHoldAnnealing
+            scheduler = NoamHoldAnnealing(optimizer, **configs['train_conf']['scheduler_conf'])
+        elif configs['train_conf']['scheduler'] == 'constantlr':
+            scheduler_type = ConstantLR
+            scheduler = ConstantLR(optimizer)
+        else:
+            raise ValueError("unknown scheduler: " + configs['train_conf'])
 
-    return model, optimizer, scheduler
+        if configs['train_conf']['optim_d'] == 'adam':
+            optimizer_d = optim.Adam(model.module.discriminator.parameters(), **configs['train_conf']['optim_conf'])
+        elif configs['train_conf']['optim_d'] == 'adamw':
+            optimizer_d = optim.AdamW(model.module.discriminator.parameters(), **configs['train_conf']['optim_conf'])
+        else:
+            raise ValueError("unknown optimizer: " + configs['train_conf'])
+
+        if configs['train_conf']['scheduler_d'] == 'warmuplr':
+            scheduler_type = WarmupLR
+            scheduler_d = WarmupLR(optimizer_d, **configs['train_conf']['scheduler_conf'])
+        elif configs['train_conf']['scheduler_d'] == 'NoamHoldAnnealing':
+            scheduler_type = NoamHoldAnnealing
+            scheduler_d = NoamHoldAnnealing(optimizer_d, **configs['train_conf']['scheduler_conf'])
+        elif configs['train_conf']['scheduler'] == 'constantlr':
+            scheduler_type = ConstantLR
+            scheduler_d = ConstantLR(optimizer_d)
+        else:
+            raise ValueError("unknown scheduler: " + configs['train_conf'])
+    return model, optimizer, scheduler, optimizer_d, scheduler_d
 
 
 def init_summarywriter(args):
@@ -157,7 +199,7 @@ def save_model(model, model_name, info_dict):
 
     if info_dict["train_engine"] == "torch_ddp":
         if rank == 0:
-            torch.save(model.module.state_dict(), save_model_path)
+            torch.save({**model.module.state_dict(), 'epoch': info_dict['epoch'], 'step': info_dict['step']}, save_model_path)
     else:
         with torch.no_grad():
             model.save_checkpoint(save_dir=model_dir,
@@ -193,7 +235,7 @@ def cosyvoice_join(group_join, info_dict):
         return False
 
 
-def batch_forward(model, batch, info_dict):
+def batch_forward(model, batch, scaler, info_dict):
     device = int(os.environ.get('LOCAL_RANK', 0))
 
     dtype = info_dict["dtype"]
@@ -205,7 +247,7 @@ def batch_forward(model, batch, info_dict):
         dtype = torch.float32
 
     if info_dict['train_engine'] == 'torch_ddp':
-        autocast = nullcontext()
+        autocast = torch.cuda.amp.autocast(enabled=scaler is not None)
     else:
         autocast = torch.cuda.amp.autocast(enabled=True, dtype=dtype, cache_enabled=False)
 
@@ -214,27 +256,41 @@ def batch_forward(model, batch, info_dict):
     return info_dict
 
 
-def batch_backward(model, info_dict):
+def batch_backward(model, scaler, info_dict):
     if info_dict["train_engine"] == "deepspeed":
         scaled_loss = model.backward(info_dict['loss_dict']['loss'])
     else:
         scaled_loss = info_dict['loss_dict']['loss'] / info_dict['accum_grad']
-        scaled_loss.backward()
+        if scaler is not None:
+            scaler.scale(scaled_loss).backward()
+        else:
+            scaled_loss.backward()
 
     info_dict['loss_dict']['loss'] = scaled_loss
     return info_dict
 
 
-def update_parameter_and_lr(model, optimizer, scheduler, info_dict):
+def update_parameter_and_lr(model, optimizer, scheduler, scaler, info_dict):
     grad_norm = 0.0
     if info_dict['train_engine'] == "deepspeed":
         info_dict["is_gradient_accumulation_boundary"] = model.is_gradient_accumulation_boundary()
         model.step()
         grad_norm = model.get_global_grad_norm()
     elif (info_dict['batch_idx'] + 1) % info_dict["accum_grad"] == 0:
-        grad_norm = clip_grad_norm_(model.parameters(), info_dict['grad_clip'])
-        if torch.isfinite(grad_norm):
-            optimizer.step()
+        # Use mixed precision training
+        if scaler is not None:
+            scaler.unscale_(optimizer)
+            grad_norm = clip_grad_norm_(model.parameters(), info_dict['grad_clip'])
+            # We don't check grad here since that if the gradient
+            # has inf/nan values, scaler.step will skip
+            # optimizer.step().
+            if torch.isfinite(grad_norm):
+                scaler.step(optimizer)
+            scaler.update()
+        else:
+            grad_norm = clip_grad_norm_(model.parameters(), info_dict['grad_clip'])
+            if torch.isfinite(grad_norm):
+                optimizer.step()
         optimizer.zero_grad()
         scheduler.step()
     info_dict["lr"] = optimizer.param_groups[0]['lr']

xinference/thirdparty/f5_tts/api.py

@@ -0,0 +1,166 @@
+import random
+import sys
+from importlib.resources import files
+
+import soundfile as sf
+import tqdm
+from cached_path import cached_path
+
+from f5_tts.infer.utils_infer import (
+    hop_length,
+    infer_process,
+    load_model,
+    load_vocoder,
+    preprocess_ref_audio_text,
+    remove_silence_for_generated_wav,
+    save_spectrogram,
+    transcribe,
+    target_sample_rate,
+)
+from f5_tts.model import DiT, UNetT
+from f5_tts.model.utils import seed_everything
+
+
+class F5TTS:
+    def __init__(
+        self,
+        model_type="F5-TTS",
+        ckpt_file="",
+        vocab_file="",
+        ode_method="euler",
+        use_ema=True,
+        vocoder_name="vocos",
+        local_path=None,
+        device=None,
+        hf_cache_dir=None,
+    ):
+        # Initialize parameters
+        self.final_wave = None
+        self.target_sample_rate = target_sample_rate
+        self.hop_length = hop_length
+        self.seed = -1
+        self.mel_spec_type = vocoder_name
+
+        # Set device
+        if device is not None:
+            self.device = device
+        else:
+            import torch
+
+            self.device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
+
+        # Load models
+        self.load_vocoder_model(vocoder_name, local_path=local_path, hf_cache_dir=hf_cache_dir)
+        self.load_ema_model(
+            model_type, ckpt_file, vocoder_name, vocab_file, ode_method, use_ema, hf_cache_dir=hf_cache_dir
+        )
+
+    def load_vocoder_model(self, vocoder_name, local_path=None, hf_cache_dir=None):
+        self.vocoder = load_vocoder(vocoder_name, local_path is not None, local_path, self.device, hf_cache_dir)
+
+    def load_ema_model(self, model_type, ckpt_file, mel_spec_type, vocab_file, ode_method, use_ema, hf_cache_dir=None):
+        if model_type == "F5-TTS":
+            if not ckpt_file:
+                if mel_spec_type == "vocos":
+                    ckpt_file = str(
+                        cached_path("hf://SWivid/F5-TTS/F5TTS_Base/model_1200000.safetensors", cache_dir=hf_cache_dir)
+                    )
+                elif mel_spec_type == "bigvgan":
+                    ckpt_file = str(
+                        cached_path("hf://SWivid/F5-TTS/F5TTS_Base_bigvgan/model_1250000.pt", cache_dir=hf_cache_dir)
+                    )
+            model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
+            model_cls = DiT
+        elif model_type == "E2-TTS":
+            if not ckpt_file:
+                ckpt_file = str(
+                    cached_path("hf://SWivid/E2-TTS/E2TTS_Base/model_1200000.safetensors", cache_dir=hf_cache_dir)
+                )
+            model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
+            model_cls = UNetT
+        else:
+            raise ValueError(f"Unknown model type: {model_type}")
+
+        self.ema_model = load_model(
+            model_cls, model_cfg, ckpt_file, mel_spec_type, vocab_file, ode_method, use_ema, self.device
+        )
+
+    def transcribe(self, ref_audio, language=None):
+        return transcribe(ref_audio, language)
+
+    def export_wav(self, wav, file_wave, remove_silence=False):
+        sf.write(file_wave, wav, self.target_sample_rate)
+
+        if remove_silence:
+            remove_silence_for_generated_wav(file_wave)
+
+    def export_spectrogram(self, spect, file_spect):
+        save_spectrogram(spect, file_spect)
+
+    def infer(
+        self,
+        ref_file,
+        ref_text,
+        gen_text,
+        show_info=print,
+        progress=tqdm,
+        target_rms=0.1,
+        cross_fade_duration=0.15,
+        sway_sampling_coef=-1,
+        cfg_strength=2,
+        nfe_step=32,
+        speed=1.0,
+        fix_duration=None,
+        remove_silence=False,
+        file_wave=None,
+        file_spect=None,
+        seed=-1,
+    ):
+        if seed == -1:
+            seed = random.randint(0, sys.maxsize)
+        seed_everything(seed)
+        self.seed = seed
+
+        ref_file, ref_text = preprocess_ref_audio_text(ref_file, ref_text, device=self.device)
+
+        wav, sr, spect = infer_process(
+            ref_file,
+            ref_text,
+            gen_text,
+            self.ema_model,
+            self.vocoder,
+            self.mel_spec_type,
+            show_info=show_info,
+            progress=progress,
+            target_rms=target_rms,
+            cross_fade_duration=cross_fade_duration,
+            nfe_step=nfe_step,
+            cfg_strength=cfg_strength,
+            sway_sampling_coef=sway_sampling_coef,
+            speed=speed,
+            fix_duration=fix_duration,
+            device=self.device,
+        )
+
+        if file_wave is not None:
+            self.export_wav(wav, file_wave, remove_silence)
+
+        if file_spect is not None:
+            self.export_spectrogram(spect, file_spect)
+
+        return wav, sr, spect
+
+
+if __name__ == "__main__":
+    f5tts = F5TTS()
+
+    wav, sr, spect = f5tts.infer(
+        ref_file=str(files("f5_tts").joinpath("infer/examples/basic/basic_ref_en.wav")),
+        ref_text="some call me nature, others call me mother nature.",
+        gen_text="""I don't really care what you call me. I've been a silent spectator, watching species evolve, empires rise and fall. But always remember, I am mighty and enduring. Respect me and I'll nurture you; ignore me and you shall face the consequences.""",
+        file_wave=str(files("f5_tts").joinpath("../../tests/api_out.wav")),
+        file_spect=str(files("f5_tts").joinpath("../../tests/api_out.png")),
+        seed=-1,  # random seed = -1
+    )
+
+    print("seed :", f5tts.seed)