update Bigvgan vocoder and F5-bigvgan version, trained on Emilia ZH&EN, 1.25m updates

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "src/third_party/BigVGAN"]
+	path = src/third_party/BigVGAN
+	url = https://github.com/NVIDIA/BigVGAN.git

README.md

@@ -46,7 +46,18 @@ cd F5-TTS
 pip install -e .
 ```
 
-### 3. Docker usage
+### 3. Init submodule( optional, if you want to change the vocoder from vocos to bigvgan)
+
+```bash
+git submodule update --init --recursive
+```
+After that, you need to change the `src/third_party/BigVGAN/bigvgan.py` by adding the following code at the beginning of the file.
+```python
+import sys
+sys.path.append(os.path.dirname(os.path.abspath(__file__)))
+```
+
+### 4. Docker usage
 ```bash
 # Build from Dockerfile
 docker build -t f5tts:v1 .

src/f5_tts/api.py

@@ -1,24 +1,18 @@
 import random
 import sys
-import tqdm
 from importlib.resources import files
 
 import soundfile as sf
 import torch
+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, target_sample_rate)
 from f5_tts.model import DiT, UNetT
 from f5_tts.model.utils import seed_everything
-from f5_tts.infer.utils_infer import (
-    load_vocoder,
-    load_model,
-    infer_process,
-    remove_silence_for_generated_wav,
-    save_spectrogram,
-    preprocess_ref_audio_text,
-    target_sample_rate,
-    hop_length,
-)
 
 
 class F5TTS:
@@ -29,6 +23,7 @@ class F5TTS:
         vocab_file="",
         ode_method="euler",
         use_ema=True,
+        vocoder_name="vocos",
         local_path=None,
         device=None,
     ):
@@ -44,11 +39,11 @@ class F5TTS:
         )
 
         # Load models
-        self.load_vocoder_model(local_path)
+        self.load_vocoder_model(vocoder_name, local_path)
         self.load_ema_model(model_type, ckpt_file, vocab_file, ode_method, use_ema)
 
-    def load_vocoder_model(self, local_path):
-        self.vocoder = load_vocoder(local_path is not None, local_path, self.device)
+    def load_vocoder_model(self, vocoder_name, local_path):
+        self.vocoder = load_vocoder(vocoder_name, local_path is not None, local_path, self.device)
 
     def load_ema_model(self, model_type, ckpt_file, vocab_file, ode_method, use_ema):
         if model_type == "F5-TTS":

src/f5_tts/eval/eval_infer_batch.py

@@ -1,26 +1,23 @@
-import sys
 import os
+import sys
 
 sys.path.append(os.getcwd())
 
-import time
-from tqdm import tqdm
 import argparse
+import time
 from importlib.resources import files
 
 import torch
 import torchaudio
 from accelerate import Accelerator
-from vocos import Vocos
+from tqdm import tqdm
 
-from f5_tts.model import CFM, UNetT, DiT
+from f5_tts.eval.utils_eval import (get_inference_prompt,
+                                    get_librispeech_test_clean_metainfo,
+                                    get_seedtts_testset_metainfo)
+from f5_tts.infer.utils_infer import load_checkpoint, load_vocoder
+from f5_tts.model import CFM, DiT, UNetT
 from f5_tts.model.utils import get_tokenizer
-from f5_tts.infer.utils_infer import load_checkpoint
-from f5_tts.eval.utils_eval import (
-    get_seedtts_testset_metainfo,
-    get_librispeech_test_clean_metainfo,
-    get_inference_prompt,
-)
 
 accelerator = Accelerator()
 device = f"cuda:{accelerator.process_index}"
@@ -31,8 +28,12 @@ device = f"cuda:{accelerator.process_index}"
 target_sample_rate = 24000
 n_mel_channels = 100
 hop_length = 256
+win_length = 1024
+n_fft = 1024
+extract_backend = "bigvgan"  # 'vocos' or 'bigvgan'
 target_rms = 0.1
 
+
 tokenizer = "pinyin"
 rel_path = str(files("f5_tts").joinpath("../../"))
 
@@ -123,14 +124,11 @@ def main():
 
     # Vocoder model
     local = False
-    if local:
-        vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
-        vocos = Vocos.from_hparams(f"{vocos_local_path}/config.yaml")
-        state_dict = torch.load(f"{vocos_local_path}/pytorch_model.bin", weights_only=True, map_location=device)
-        vocos.load_state_dict(state_dict)
-        vocos.eval()
-    else:
-        vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
+    if extract_backend == "vocos":
+        vocoder_local_path = "../checkpoints/charactr/vocos-mel-24khz"
+    elif extract_backend == "bigvgan":
+        vocoder_local_path = "../checkpoints/bigvgan_v2_24khz_100band_256x"
+    vocoder = load_vocoder(vocoder_name=extract_backend, is_local=local, local_path=vocoder_local_path)
 
     # Tokenizer
     vocab_char_map, vocab_size = get_tokenizer(dataset_name, tokenizer)
@@ -139,9 +137,12 @@ def main():
     model = CFM(
         transformer=model_cls(**model_cfg, text_num_embeds=vocab_size, mel_dim=n_mel_channels),
         mel_spec_kwargs=dict(
-            target_sample_rate=target_sample_rate,
-            n_mel_channels=n_mel_channels,
+            n_fft=n_fft,
             hop_length=hop_length,
+            win_length=win_length,
+            n_mel_channels=n_mel_channels,
+            target_sample_rate=target_sample_rate,
+            extract_backend=extract_backend,
         ),
         odeint_kwargs=dict(
             method=ode_method,
@@ -149,7 +150,8 @@ def main():
         vocab_char_map=vocab_char_map,
     ).to(device)
 
-    model = load_checkpoint(model, ckpt_path, device, use_ema=use_ema)
+    dtype = torch.float16 if extract_backend == "vocos" else torch.float32
+    model = load_checkpoint(model, ckpt_path, device, dtype, use_ema=use_ema)
 
     if not os.path.exists(output_dir) and accelerator.is_main_process:
         os.makedirs(output_dir)
@@ -178,14 +180,18 @@ def main():
                     no_ref_audio=no_ref_audio,
                     seed=seed,
                 )
-            # Final result
-            for i, gen in enumerate(generated):
-                gen = gen[ref_mel_lens[i] : total_mel_lens[i], :].unsqueeze(0)
-                gen_mel_spec = gen.permute(0, 2, 1)
-                generated_wave = vocos.decode(gen_mel_spec.cpu())
-                if ref_rms_list[i] < target_rms:
-                    generated_wave = generated_wave * ref_rms_list[i] / target_rms
-                torchaudio.save(f"{output_dir}/{utts[i]}.wav", generated_wave, target_sample_rate)
+                # Final result
+                for i, gen in enumerate(generated):
+                    gen = gen[ref_mel_lens[i] : total_mel_lens[i], :].unsqueeze(0)
+                    gen_mel_spec = gen.permute(0, 2, 1)
+                    if extract_backend == "vocos":
+                        generated_wave = vocoder.decode(gen_mel_spec.cpu())
+                    elif extract_backend == "bigvgan":
+                        generated_wave = vocoder(gen_mel_spec)
+
+                    if ref_rms_list[i] < target_rms:
+                        generated_wave = generated_wave * ref_rms_list[i] / target_rms
+                    torchaudio.save(f"{output_dir}/{utts[i]}.wav", generated_wave.squeeze(0).cpu(), target_sample_rate)
 
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:

src/f5_tts/eval/utils_eval.py

@@ -2,15 +2,15 @@ import math
 import os
 import random
 import string
-from tqdm import tqdm
 
 import torch
 import torch.nn.functional as F
 import torchaudio
+from tqdm import tqdm
 
+from f5_tts.eval.ecapa_tdnn import ECAPA_TDNN_SMALL
 from f5_tts.model.modules import MelSpec
 from f5_tts.model.utils import convert_char_to_pinyin
-from f5_tts.eval.ecapa_tdnn import ECAPA_TDNN_SMALL
 
 
 # seedtts testset metainfo: utt, prompt_text, prompt_wav, gt_text, gt_wav
@@ -74,8 +74,11 @@ def get_inference_prompt(
     tokenizer="pinyin",
     polyphone=True,
     target_sample_rate=24000,
+    n_fft=1024,
+    win_length=1024,
     n_mel_channels=100,
     hop_length=256,
+    extract_backend="bigvgan",
     target_rms=0.1,
     use_truth_duration=False,
     infer_batch_size=1,
@@ -94,7 +97,12 @@ def get_inference_prompt(
     )
 
     mel_spectrogram = MelSpec(
-        target_sample_rate=target_sample_rate, n_mel_channels=n_mel_channels, hop_length=hop_length
+        n_fft=n_fft,
+        hop_length=hop_length,
+        win_length=win_length,
+        n_mel_channels=n_mel_channels,
+        target_sample_rate=target_sample_rate,
+        extract_backend=extract_backend,
     )
 
     for utt, prompt_text, prompt_wav, gt_text, gt_wav in tqdm(metainfo, desc="Processing prompts..."):

src/f5_tts/infer/infer_cli.py

@@ -2,23 +2,18 @@ import argparse
 import codecs
 import os
 import re
-from pathlib import Path
 from importlib.resources import files
+from pathlib import Path
 
 import numpy as np
 import soundfile as sf
 import tomli
 from cached_path import cached_path
 
+from f5_tts.infer.utils_infer import (infer_process, load_model, load_vocoder,
+                                      preprocess_ref_audio_text,
+                                      remove_silence_for_generated_wav)
 from f5_tts.model import DiT, UNetT
-from f5_tts.infer.utils_infer import (
-    load_vocoder,
-    load_model,
-    preprocess_ref_audio_text,
-    infer_process,
-    remove_silence_for_generated_wav,
-)
-
 
 parser = argparse.ArgumentParser(
     prog="python3 infer-cli.py",
@@ -70,6 +65,7 @@ parser.add_argument(
     "--remove_silence",
     help="Remove silence.",
 )
+parser.add_argument("--vocoder_name", type=str, default="vocos", choices=["vocos", "bigvgan"], help="vocoder name")
 parser.add_argument(
     "--load_vocoder_from_local",
     action="store_true",
@@ -111,9 +107,14 @@ remove_silence = args.remove_silence if args.remove_silence else config["remove_
 speed = args.speed
 wave_path = Path(output_dir) / "infer_cli_out.wav"
 # spectrogram_path = Path(output_dir) / "infer_cli_out.png"
-vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
+if args.vocoder_name == "vocos":
+    vocoder_local_path = "../checkpoints/charactr/vocos-mel-24khz"
+elif args.vocoder_name == "bigvgan":
+    vocoder_local_path = "../checkpoints/bigvgan_v2_24khz_100band_256x"
 
-vocoder = load_vocoder(is_local=args.load_vocoder_from_local, local_path=vocos_local_path)
+vocoder = load_vocoder(
+    vocoder_name=args.vocoder_name, is_local=args.load_vocoder_from_local, local_path=vocoder_local_path
+)
 
 
 # load models
@@ -136,6 +137,12 @@ elif model == "E2-TTS":
         ckpt_step = 1200000
         ckpt_file = str(cached_path(f"hf://SWivid/{repo_name}/{exp_name}/model_{ckpt_step}.safetensors"))
         # ckpt_file = f"ckpts/{exp_name}/model_{ckpt_step}.pt"  # .pt | .safetensors; local path
+    elif args.vocoder_name == "bigvgan":  # TODO: need to test
+        repo_name = "F5-TTS"
+        exp_name = "F5TTS_Base_bigvgan"
+        ckpt_step = 1250000
+        ckpt_file = str(cached_path(f"hf://SWivid/{repo_name}/{exp_name}/model_{ckpt_step}.pt"))
+
 
 print(f"Using {model}...")
 ema_model = load_model(model_cls, model_cfg, ckpt_file, vocab_file)

src/f5_tts/infer/speech_edit.py

@@ -3,17 +3,11 @@ import os
 import torch
 import torch.nn.functional as F
 import torchaudio
-from vocos import Vocos
-
-from f5_tts.model import CFM, UNetT, DiT
-from f5_tts.model.utils import (
-    get_tokenizer,
-    convert_char_to_pinyin,
-)
-from f5_tts.infer.utils_infer import (
-    load_checkpoint,
-    save_spectrogram,
-)
+
+from f5_tts.infer.utils_infer import (load_checkpoint, load_vocoder,
+                                      save_spectrogram)
+from f5_tts.model import CFM, DiT, UNetT
+from f5_tts.model.utils import convert_char_to_pinyin, get_tokenizer
 
 device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
 
@@ -23,6 +17,9 @@ device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is
 target_sample_rate = 24000
 n_mel_channels = 100
 hop_length = 256
+win_length = 1024
+n_fft = 1024
+extract_backend = "bigvgan"  # 'vocos' or 'bigvgan'
 target_rms = 0.1
 
 tokenizer = "pinyin"
@@ -89,15 +86,11 @@ if not os.path.exists(output_dir):
 
 # Vocoder model
 local = False
-if local:
-    vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
-    vocos = Vocos.from_hparams(f"{vocos_local_path}/config.yaml")
-    state_dict = torch.load(f"{vocos_local_path}/pytorch_model.bin", weights_only=True, map_location=device)
-    vocos.load_state_dict(state_dict)
-
-    vocos.eval()
-else:
-    vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
+if extract_backend == "vocos":
+    vocoder_local_path = "../checkpoints/charactr/vocos-mel-24khz"
+elif extract_backend == "bigvgan":
+    vocoder_local_path = "../checkpoints/bigvgan_v2_24khz_100band_256x"
+vocoder = load_vocoder(vocoder_name=extract_backend, is_local=local, local_path=vocoder_local_path)
 
 # Tokenizer
 vocab_char_map, vocab_size = get_tokenizer(dataset_name, tokenizer)
@@ -106,9 +99,12 @@ vocab_char_map, vocab_size = get_tokenizer(dataset_name, tokenizer)
 model = CFM(
     transformer=model_cls(**model_cfg, text_num_embeds=vocab_size, mel_dim=n_mel_channels),
     mel_spec_kwargs=dict(
-        target_sample_rate=target_sample_rate,
-        n_mel_channels=n_mel_channels,
+        n_fft=n_fft,
         hop_length=hop_length,
+        win_length=win_length,
+        n_mel_channels=n_mel_channels,
+        target_sample_rate=target_sample_rate,
+        extract_backend=extract_backend,
     ),
     odeint_kwargs=dict(
         method=ode_method,
@@ -116,7 +112,8 @@ model = CFM(
     vocab_char_map=vocab_char_map,
 ).to(device)
 
-model = load_checkpoint(model, ckpt_path, device, use_ema=use_ema)
+dtype = torch.float16 if extract_backend == "vocos" else torch.float32
+model = load_checkpoint(model, ckpt_path, device, dtype, use_ema=use_ema)
 
 # Audio
 audio, sr = torchaudio.load(audio_to_edit)
@@ -181,11 +178,15 @@ print(f"Generated mel: {generated.shape}")
 # Final result
 generated = generated.to(torch.float32)
 generated = generated[:, ref_audio_len:, :]
-generated_mel_spec = generated.permute(0, 2, 1)
-generated_wave = vocos.decode(generated_mel_spec.cpu())
+gen_mel_spec = generated.permute(0, 2, 1)
+if extract_backend == "vocos":
+    generated_wave = vocoder.decode(gen_mel_spec.cpu())
+elif extract_backend == "bigvgan":
+    generated_wave = vocoder(gen_mel_spec)
+
 if rms < target_rms:
     generated_wave = generated_wave * rms / target_rms
 
-save_spectrogram(generated_mel_spec[0].cpu().numpy(), f"{output_dir}/speech_edit_out.png")
-torchaudio.save(f"{output_dir}/speech_edit_out.wav", generated_wave, target_sample_rate)
+save_spectrogram(gen_mel_spec[0].cpu().numpy(), f"{output_dir}/speech_edit_out.png")
+torchaudio.save(f"{output_dir}/speech_edit_out.wav", generated_wave.squeeze(0).cpu(), target_sample_rate)
 print(f"Generated wav: {generated_wave.shape}")

src/f5_tts/infer/utils_infer.py

@@ -1,6 +1,10 @@
 # A unified script for inference process
 # Make adjustments inside functions, and consider both gradio and cli scripts if need to change func output format
+import os
+import sys
 
+sys.path.append(f"../../{os.path.dirname(os.path.abspath(__file__))}/third_party/BigVGAN/")
+from third_party.BigVGAN import bigvgan
 import hashlib
 import re
 import tempfile
@@ -34,6 +38,9 @@ device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is
 target_sample_rate = 24000
 n_mel_channels = 100
 hop_length = 256
+win_length = 1024
+n_fft = 1024
+extract_backend = "bigvgan"  # 'vocos' or 'bigvgan'
 target_rms = 0.1
 cross_fade_duration = 0.15
 ode_method = "euler"
@@ -80,17 +87,28 @@ def chunk_text(text, max_chars=135):
 
 
 # load vocoder
-def load_vocoder(is_local=False, local_path="", device=device):
-    if is_local:
-        print(f"Load vocos from local path {local_path}")
-        vocos = Vocos.from_hparams(f"{local_path}/config.yaml")
-        state_dict = torch.load(f"{local_path}/pytorch_model.bin", map_location=device)
-        vocos.load_state_dict(state_dict)
-        vocos.eval()
-    else:
-        print("Download Vocos from huggingface charactr/vocos-mel-24khz")
-        vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
-    return vocos
+def load_vocoder(vocoder_name="vocos", is_local=False, local_path="", device=device):
+    if vocoder_name == "vocos":
+        if is_local:
+            print(f"Load vocos from local path {local_path}")
+            vocoder = Vocos.from_hparams(f"{local_path}/config.yaml")
+            state_dict = torch.load(f"{local_path}/pytorch_model.bin", map_location="cpu")
+            vocoder.load_state_dict(state_dict)
+            vocoder.eval()
+            vocoder = vocoder.eval().to(device)
+        else:
+            print("Download Vocos from huggingface charactr/vocos-mel-24khz")
+            vocoder = Vocos.from_pretrained("charactr/vocos-mel-24khz")
+    elif vocoder_name == "bigvgan":
+        if is_local:
+            """download from https://huggingface.co/nvidia/bigvgan_v2_24khz_100band_256x/tree/main"""
+            vocoder = bigvgan.BigVGAN.from_pretrained(local_path, use_cuda_kernel=False)
+        else:
+            vocoder = bigvgan.BigVGAN.from_pretrained("nvidia/bigvgan_v2_24khz_100band_256x", use_cuda_kernel=False)
+
+        vocoder.remove_weight_norm()
+        vocoder = vocoder.eval().to(device)
+    return vocoder
 
 
 # load asr pipeline
@@ -111,9 +129,8 @@ def initialize_asr_pipeline(device=device):
 # load model checkpoint for inference
 
 
-def load_checkpoint(model, ckpt_path, device, use_ema=True):
-    if device == "cuda":
-        model = model.half()
+def load_checkpoint(model, ckpt_path, device, dtype, use_ema=True):
+    model = model.to(dtype)
 
     ckpt_type = ckpt_path.split(".")[-1]
     if ckpt_type == "safetensors":
@@ -156,9 +173,12 @@ def load_model(model_cls, model_cfg, ckpt_path, vocab_file="", ode_method=ode_me
     model = CFM(
         transformer=model_cls(**model_cfg, text_num_embeds=vocab_size, mel_dim=n_mel_channels),
         mel_spec_kwargs=dict(
-            target_sample_rate=target_sample_rate,
-            n_mel_channels=n_mel_channels,
+            n_fft=n_fft,
             hop_length=hop_length,
+            win_length=win_length,
+            n_mel_channels=n_mel_channels,
+            target_sample_rate=target_sample_rate,
+            extract_backend=extract_backend,
         ),
         odeint_kwargs=dict(
             method=ode_method,
@@ -166,7 +186,8 @@ def load_model(model_cls, model_cfg, ckpt_path, vocab_file="", ode_method=ode_me
         vocab_char_map=vocab_char_map,
     ).to(device)
 
-    model = load_checkpoint(model, ckpt_path, device, use_ema=use_ema)
+    dtype = torch.float16 if extract_backend == "vocos" else torch.float32
+    model = load_checkpoint(model, ckpt_path, device, dtype, use_ema=use_ema)
 
     return model
 
@@ -359,18 +380,21 @@ def infer_batch_process(
                 sway_sampling_coef=sway_sampling_coef,
             )
 
-        generated = generated.to(torch.float32)
-        generated = generated[:, ref_audio_len:, :]
-        generated_mel_spec = generated.permute(0, 2, 1)
-        generated_wave = vocoder.decode(generated_mel_spec.cpu())
-        if rms < target_rms:
-            generated_wave = generated_wave * rms / target_rms
-
-        # wav -> numpy
-        generated_wave = generated_wave.squeeze().cpu().numpy()
-
-        generated_waves.append(generated_wave)
-        spectrograms.append(generated_mel_spec[0].cpu().numpy())
+            generated = generated.to(torch.float32)
+            generated = generated[:, ref_audio_len:, :]
+            generated_mel_spec = generated.permute(0, 2, 1)
+            if extract_backend == "vocos":
+                generated_wave = vocoder.decode(generated_mel_spec.cpu())
+            elif extract_backend == "bigvgan":
+                generated_wave = vocoder(generated_mel_spec)
+            if rms < target_rms:
+                generated_wave = generated_wave * rms / target_rms
+
+            # wav -> numpy
+            generated_wave = generated_wave.squeeze().cpu().numpy()
+
+            generated_waves.append(generated_wave)
+            spectrograms.append(generated_mel_spec[0].cpu().numpy())
 
     # Combine all generated waves with cross-fading
     if cross_fade_duration <= 0:

src/f5_tts/model/cfm.py

@@ -8,25 +8,19 @@ d - dimension
 """
 
 from __future__ import annotations
-from typing import Callable
+
 from random import random
+from typing import Callable
 
 import torch
-from torch import nn
 import torch.nn.functional as F
+from torch import nn
 from torch.nn.utils.rnn import pad_sequence
-
 from torchdiffeq import odeint
 
 from f5_tts.model.modules import MelSpec
-from f5_tts.model.utils import (
-    default,
-    exists,
-    list_str_to_idx,
-    list_str_to_tensor,
-    lens_to_mask,
-    mask_from_frac_lengths,
-)
+from f5_tts.model.utils import (default, exists, lens_to_mask, list_str_to_idx,
+                                list_str_to_tensor, mask_from_frac_lengths)
 
 
 class CFM(nn.Module):
@@ -99,8 +93,10 @@ class CFM(nn.Module):
     ):
         self.eval()
 
-        if next(self.parameters()).dtype == torch.float16:
-            cond = cond.half()
+        assert next(self.parameters()).dtype == torch.float32 or next(self.parameters()).dtype == torch.float16, print(
+            "Only support fp16 and fp32 inference currently"
+        )
+        cond = cond.to(next(self.parameters()).dtype)
 
         # raw wave
 

src/f5_tts/model/dataset.py

@@ -1,15 +1,15 @@
 import json
 import random
 from importlib.resources import files
-from tqdm import tqdm
 
 import torch
 import torch.nn.functional as F
 import torchaudio
+from datasets import Dataset as Dataset_
+from datasets import load_from_disk
 from torch import nn
 from torch.utils.data import Dataset, Sampler
-from datasets import load_from_disk
-from datasets import Dataset as Dataset_
+from tqdm import tqdm
 
 from f5_tts.model.modules import MelSpec
 from f5_tts.model.utils import default
@@ -22,12 +22,21 @@ class HFDataset(Dataset):
         target_sample_rate=24_000,
         n_mel_channels=100,
         hop_length=256,
+        n_fft=1024,
+        win_length=1024,
+        extract_backend="vocos",
     ):
         self.data = hf_dataset
         self.target_sample_rate = target_sample_rate
         self.hop_length = hop_length
+
         self.mel_spectrogram = MelSpec(
-            target_sample_rate=target_sample_rate, n_mel_channels=n_mel_channels, hop_length=hop_length
+            n_fft=n_fft,
+            hop_length=hop_length,
+            win_length=win_length,
+            n_mel_channels=n_mel_channels,
+            target_sample_rate=target_sample_rate,
+            extract_backend=extract_backend,
         )
 
     def get_frame_len(self, index):
@@ -79,6 +88,9 @@ class CustomDataset(Dataset):
         target_sample_rate=24_000,
         hop_length=256,
         n_mel_channels=100,
+        n_fft=1024,
+        win_length=1024,
+        extract_backend="vocos",
         preprocessed_mel=False,
         mel_spec_module: nn.Module | None = None,
     ):
@@ -86,15 +98,21 @@ class CustomDataset(Dataset):
         self.durations = durations
         self.target_sample_rate = target_sample_rate
         self.hop_length = hop_length
+        self.n_fft = n_fft
+        self.win_length = win_length
+        self.extract_backend = extract_backend
         self.preprocessed_mel = preprocessed_mel
 
         if not preprocessed_mel:
             self.mel_spectrogram = default(
                 mel_spec_module,
                 MelSpec(
-                    target_sample_rate=target_sample_rate,
+                    n_fft=n_fft,
                     hop_length=hop_length,
+                    win_length=win_length,
                     n_mel_channels=n_mel_channels,
+                    target_sample_rate=target_sample_rate,
+                    extract_backend=extract_backend,
                 ),
             )
 

src/f5_tts/model/modules.py

@@ -8,61 +8,173 @@ d - dimension
 """
 
 from __future__ import annotations
-from typing import Optional
+
 import math
+from typing import Optional
 
 import torch
-from torch import nn
 import torch.nn.functional as F
 import torchaudio
-
+from librosa.filters import mel as librosa_mel_fn
+from torch import nn
 from x_transformers.x_transformers import apply_rotary_pos_emb
 
-
 # raw wav to mel spec
 
 
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    return dynamic_range_compression_torch(magnitudes)
+
+
+mel_basis_cache = {}
+hann_window_cache = {}
+
+
+# BigVGAN extract mel spectrogram
+def mel_spectrogram(
+    y: torch.Tensor,
+    n_fft: int,
+    num_mels: int,
+    sampling_rate: int,
+    hop_size: int,
+    win_size: int,
+    fmin: int,
+    fmax: int = None,
+    center: bool = False,
+) -> torch.Tensor:
+    """Copy from https://github.com/NVIDIA/BigVGAN/tree/main"""
+    device = y.device
+    key = f"{n_fft}_{num_mels}_{sampling_rate}_{hop_size}_{win_size}_{fmin}_{fmax}_{device}"
+
+    if key not in mel_basis_cache:
+        mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
+        mel_basis_cache[key] = torch.from_numpy(mel).float().to(device)  # TODO: why they need .float()?
+        hann_window_cache[key] = torch.hann_window(win_size).to(device)
+
+    mel_basis = mel_basis_cache[key]
+    hann_window = hann_window_cache[key]
+
+    padding = (n_fft - hop_size) // 2
+    y = torch.nn.functional.pad(y.unsqueeze(1), (padding, padding), mode="reflect").squeeze(1)
+
+    spec = torch.stft(
+        y,
+        n_fft,
+        hop_length=hop_size,
+        win_length=win_size,
+        window=hann_window,
+        center=center,
+        pad_mode="reflect",
+        normalized=False,
+        onesided=True,
+        return_complex=True,
+    )
+    spec = torch.sqrt(torch.view_as_real(spec).pow(2).sum(-1) + 1e-9)
+
+    mel_spec = torch.matmul(mel_basis, spec)
+    mel_spec = spectral_normalize_torch(mel_spec)
+
+    return mel_spec
+
+
+def get_bigvgan_mel_spectrogram(
+    waveform,
+    n_fft=1024,
+    n_mel_channels=100,
+    target_sample_rate=24000,
+    hop_length=256,
+    win_length=1024,
+):
+    return mel_spectrogram(
+        waveform,
+        n_fft,  # 1024
+        n_mel_channels,  # 100
+        target_sample_rate,  # 24000
+        hop_length,  # 256
+        win_length,  # 1024
+        fmin=0,  # 0
+        fmax=None,  # null
+    )
+
+
+def get_vocos_mel_spectrogram(
+    waveform,
+    n_fft=1024,
+    n_mel_channels=100,
+    target_sample_rate=24000,
+    hop_length=256,
+    win_length=1024,
+):
+    mel_stft = torchaudio.transforms.MelSpectrogram(
+        sample_rate=target_sample_rate,
+        n_fft=n_fft,
+        win_length=win_length,
+        hop_length=hop_length,
+        n_mels=n_mel_channels,
+        power=1,
+        center=True,
+        normalized=False,
+        norm=None,
+    )
+    if len(waveform.shape) == 3:
+        waveform = waveform.squeeze(1)  # 'b 1 nw -> b nw'
+
+    assert len(waveform.shape) == 2
+
+    mel = mel_stft(waveform)
+    mel = mel.clamp(min=1e-5).log()
+    return mel
+
+
 class MelSpec(nn.Module):
     def __init__(
         self,
-        filter_length=1024,
+        n_fft=1024,
         hop_length=256,
         win_length=1024,
         n_mel_channels=100,
         target_sample_rate=24_000,
-        normalize=False,
-        power=1,
-        norm=None,
-        center=True,
+        extract_backend="vocos",
     ):
         super().__init__()
-        self.n_mel_channels = n_mel_channels
-
-        self.mel_stft = torchaudio.transforms.MelSpectrogram(
-            sample_rate=target_sample_rate,
-            n_fft=filter_length,
-            win_length=win_length,
-            hop_length=hop_length,
-            n_mels=n_mel_channels,
-            power=power,
-            center=center,
-            normalized=normalize,
-            norm=norm,
+        assert extract_backend in ["vocos", "bigvgan"], print(
+            "We only support two extract mel backend: vocos or bigvgan"
         )
 
-        self.register_buffer("dummy", torch.tensor(0), persistent=False)
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.n_mel_channels = n_mel_channels
+        self.target_sample_rate = target_sample_rate
 
-    def forward(self, inp):
-        if len(inp.shape) == 3:
-            inp = inp.squeeze(1)  # 'b 1 nw -> b nw'
+        if extract_backend == "vocos":
+            self.extractor = get_vocos_mel_spectrogram
+        elif extract_backend == "bigvgan":
+            self.extractor = get_bigvgan_mel_spectrogram
 
-        assert len(inp.shape) == 2
+        self.register_buffer("dummy", torch.tensor(0), persistent=False)
 
-        if self.dummy.device != inp.device:
-            self.to(inp.device)
+    def forward(self, wav):
+        if self.dummy.device != wav.device:
+            self.to(wav.device)
+
+        mel = self.extractor(
+            waveform=wav,
+            n_fft=self.n_fft,
+            n_mel_channels=self.n_mel_channels,
+            target_sample_rate=self.target_sample_rate,
+            hop_length=self.hop_length,
+            win_length=self.win_length,
+        )
 
-        mel = self.mel_stft(inp)
-        mel = mel.clamp(min=1e-5).log()
         return mel
 
 

src/f5_tts/model/trainer.py

@@ -1,25 +1,22 @@
 from __future__ import annotations
 
-import os
 import gc
-from tqdm import tqdm
-import wandb
+import os
 
 import torch
 import torchaudio
-from torch.optim import AdamW
-from torch.utils.data import DataLoader, Dataset, SequentialSampler
-from torch.optim.lr_scheduler import LinearLR, SequentialLR
-
+import wandb
 from accelerate import Accelerator
 from accelerate.utils import DistributedDataParallelKwargs
-
 from ema_pytorch import EMA
+from torch.optim import AdamW
+from torch.optim.lr_scheduler import LinearLR, SequentialLR
+from torch.utils.data import DataLoader, Dataset, SequentialSampler
+from tqdm import tqdm
 
 from f5_tts.model import CFM
-from f5_tts.model.utils import exists, default
 from f5_tts.model.dataset import DynamicBatchSampler, collate_fn
-
+from f5_tts.model.utils import default, exists
 
 # trainer
 
@@ -49,6 +46,7 @@ class Trainer:
         accelerate_kwargs: dict = dict(),
         ema_kwargs: dict = dict(),
         bnb_optimizer: bool = False,
+        extract_backend: str = "vocos",  # "vocos" | "bigvgan"
     ):
         ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
 
@@ -110,6 +108,7 @@ class Trainer:
         self.max_samples = max_samples
         self.grad_accumulation_steps = grad_accumulation_steps
         self.max_grad_norm = max_grad_norm
+        self.vocoder_name = extract_backend
 
         self.noise_scheduler = noise_scheduler
 
@@ -188,9 +187,10 @@ class Trainer:
 
     def train(self, train_dataset: Dataset, num_workers=16, resumable_with_seed: int = None):
         if self.log_samples:
-            from f5_tts.infer.utils_infer import load_vocoder, nfe_step, cfg_strength, sway_sampling_coef
+            from f5_tts.infer.utils_infer import (cfg_strength, load_vocoder,
+                                                  nfe_step, sway_sampling_coef)
 
-            vocoder = load_vocoder()
+            vocoder = load_vocoder(vocoder_name=self.vocoder_name)
             target_sample_rate = self.accelerator.unwrap_model(self.model).mel_spec.mel_stft.sample_rate
             log_samples_path = f"{self.checkpoint_path}/samples"
             os.makedirs(log_samples_path, exist_ok=True)

src/f5_tts/train/train.py

@@ -2,16 +2,18 @@
 
 from importlib.resources import files
 
-from f5_tts.model import CFM, UNetT, DiT, Trainer
-from f5_tts.model.utils import get_tokenizer
+from f5_tts.model import CFM, DiT, Trainer, UNetT
 from f5_tts.model.dataset import load_dataset
-
+from f5_tts.model.utils import get_tokenizer
 
 # -------------------------- Dataset Settings --------------------------- #
 
 target_sample_rate = 24000
 n_mel_channels = 100
 hop_length = 256
+win_length = 1024
+n_fft = 1024
+extract_backend = "bigvgan"  # 'vocos' or 'bigvgan'
 
 tokenizer = "pinyin"  # 'pinyin', 'char', or 'custom'
 tokenizer_path = None  # if tokenizer = 'custom', define the path to the tokenizer you want to use (should be vocab.txt)
@@ -56,9 +58,12 @@ def main():
     vocab_char_map, vocab_size = get_tokenizer(tokenizer_path, tokenizer)
 
     mel_spec_kwargs = dict(
-        target_sample_rate=target_sample_rate,
-        n_mel_channels=n_mel_channels,
+        n_fft=n_fft,
         hop_length=hop_length,
+        win_length=win_length,
+        n_mel_channels=n_mel_channels,
+        target_sample_rate=target_sample_rate,
+        extract_backend=extract_backend,
     )
 
     model = CFM(
@@ -84,6 +89,7 @@ def main():
         wandb_resume_id=wandb_resume_id,
         last_per_steps=last_per_steps,
         log_samples=True,
+        extract_backend=extract_backend,
     )
 
     train_dataset = load_dataset(dataset_name, tokenizer, mel_spec_kwargs=mel_spec_kwargs)

src/third_party/BigVGAN

@@ -0,0 +1 @@
+Subproject commit 7d2b454564a6c7d014227f635b7423881f14bdac