[Bug] XTTS Training with LJSpeech leads to weird inference output #362
Description
Describe the bug
I tried training XTTS model with LJSpeech dataset, using the xtts v2 training script from the recipes folder.
However, when doing inference, I don't get any speech, only loud metallic buzzing noise. I tried doing inference with original model and it worked fine. I also tried to train with custom tokenizer, mel_stats but the issue still persists. I will share my training script, inference script and config.json file. Thank you in advance.
To Reproduce
Training Script
import os
from trainer import Trainer, TrainerArgs
from TTS.config.shared_configs import BaseDatasetConfig
from TTS.tts.datasets import load_tts_samples
from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig
from TTS.tts.models.xtts import XttsAudioConfig
from TTS.utils.manage import ModelManager
# Logging parameters
RUN_NAME = "GPT_XTTS_v2.0_LJSpeech_FT"
PROJECT_NAME = "XTTS_trainer"
DASHBOARD_LOGGER = "tensorboard"
LOGGER_URI = None
output_path = os.path.dirname(os.path.abspath(__file__))
# Set here the path that the checkpoints will be saved. Default: ./run/training/
OUT_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), "run", "training")
# Training Parameters
OPTIMIZER_WD_ONLY_ON_WEIGHTS = True # for multi-gpu training please make it False
START_WITH_EVAL = True # if True it will star with evaluation
BATCH_SIZE = 32 # set here the batch size
GRAD_ACUMM_STEPS = 84 # set here the grad accumulation steps
# Note: we recommend that BATCH_SIZE * GRAD_ACUMM_STEPS need to be at least 252 for more efficient training. You can increase/decrease BATCH_SIZE but then set GRAD_ACUMM_STEPS accordingly.
# Define here the dataset that you want to use for the fine-tuning on.
config_dataset = BaseDatasetConfig(
formatter="ljspeech",
dataset_name="ljspeech",
path=os.path.join(output_path, "../LJSpeech-1.1/"),
meta_file_train="metadata.csv",
language="en",
)
# Add here the configs of the datasets
DATASETS_CONFIG_LIST = [config_dataset]
# Define the path where XTTS v2.0.1 files will be downloaded
CHECKPOINTS_OUT_PATH = os.path.join(OUT_PATH, "XTTS_v2.0_original_model_files/")
os.makedirs(CHECKPOINTS_OUT_PATH, exist_ok=True)
# DVAE files
DVAE_CHECKPOINT_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/dvae.pth"
MEL_NORM_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/mel_stats.pth"
# Set the path to the downloaded files
DVAE_CHECKPOINT = os.path.join(CHECKPOINTS_OUT_PATH, os.path.basename(DVAE_CHECKPOINT_LINK))
MEL_NORM_FILE = os.path.join(CHECKPOINTS_OUT_PATH, os.path.basename(MEL_NORM_LINK))
# download DVAE files if needed
if not os.path.isfile(DVAE_CHECKPOINT) or not os.path.isfile(MEL_NORM_FILE):
print(" > Downloading DVAE files!")
ModelManager._download_model_files([MEL_NORM_LINK, DVAE_CHECKPOINT_LINK], CHECKPOINTS_OUT_PATH, progress_bar=True)
# Download XTTS v2.0 checkpoint if needed
TOKENIZER_FILE_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/vocab.json"
XTTS_CHECKPOINT_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/model.pth"
# XTTS transfer learning parameters: You we need to provide the paths of XTTS model checkpoint that you want to do the fine tuning.
TOKENIZER_FILE = os.path.join(CHECKPOINTS_OUT_PATH, os.path.basename(TOKENIZER_FILE_LINK)) # vocab.json file
XTTS_CHECKPOINT = os.path.join(CHECKPOINTS_OUT_PATH, os.path.basename(XTTS_CHECKPOINT_LINK)) # model.pth file
# download XTTS v2.0 files if needed
if not os.path.isfile(TOKENIZER_FILE) or not os.path.isfile(XTTS_CHECKPOINT):
print(" > Downloading XTTS v2.0 files!")
ModelManager._download_model_files(
[TOKENIZER_FILE_LINK, XTTS_CHECKPOINT_LINK], CHECKPOINTS_OUT_PATH, progress_bar=True
)
# Training sentences generations
SPEAKER_REFERENCE = [
os.path.join(output_path, "../LJSpeech-1.1/wavs/LJ001-0002.wav") # speaker reference to be used in training test sentences
]
LANGUAGE = config_dataset.language
TOKENIZER_FILE_LJS = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tokenizer_ljs.json')
MEL_NORM_FILE_LJS = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'mel_mean_ljs.pth')
XTTS_CHECKPOINT = os.path.join(OUT_PATH, 'GPT_XTTS_v2.0_LJSpeech_FT-April-02-2025_09+37PM-0000000/model.pth')
def main():
# init args and config
model_args = GPTArgs(
max_conditioning_length=132300, # 6 secs
min_conditioning_length=66150, # 3 secs
debug_loading_failures=False,
max_wav_length=255995, # ~11.6 seconds
max_text_length=200,
mel_norm_file=MEL_NORM_FILE_LJS,
dvae_checkpoint=DVAE_CHECKPOINT,
xtts_checkpoint=XTTS_CHECKPOINT, # checkpoint path of the model that you want to fine-tune
tokenizer_file=TOKENIZER_FILE_LJS,
gpt_num_audio_tokens=1026,
gpt_start_audio_token=1024,
gpt_stop_audio_token=1025,
gpt_use_masking_gt_prompt_approach=True,
gpt_use_perceiver_resampler=True,
)
# define audio config
audio_config = XttsAudioConfig(sample_rate=22050, dvae_sample_rate=22050, output_sample_rate=24000)
# training parameters config
config = GPTTrainerConfig(
output_path=OUT_PATH,
model_args=model_args,
run_name=RUN_NAME,
project_name=PROJECT_NAME,
run_description="""
GPT XTTS training
""",
dashboard_logger=DASHBOARD_LOGGER,
logger_uri=LOGGER_URI,
audio=audio_config,
batch_size=BATCH_SIZE,
batch_group_size=48,
eval_batch_size=int(BATCH_SIZE / 2),
num_loader_workers=4,
eval_split_max_size=256,
print_step=50,
plot_step=100,
log_model_step=1000,
save_step=10000,
save_n_checkpoints=0,
save_checkpoints=False,
# target_loss="loss",
print_eval=False,
# Optimizer values like tortoise, pytorch implementation with modifications to not apply WD to non-weight parameters.
optimizer="AdamW",
optimizer_wd_only_on_weights=OPTIMIZER_WD_ONLY_ON_WEIGHTS,
optimizer_params={"betas": [0.9, 0.96], "eps": 1e-8, "weight_decay": 1e-2},
lr=5e-06, # learning rate
lr_scheduler="MultiStepLR",
# it was adjusted accordly for the new step scheme
lr_scheduler_params={"milestones": [50000 * 18, 150000 * 18, 300000 * 18], "gamma": 0.5, "last_epoch": -1},
test_sentences=[
{
"text": "It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.",
"speaker_wav": SPEAKER_REFERENCE,
"language": LANGUAGE,
},
{
"text": "This cake is great. It's so delicious and moist.",
"speaker_wav": SPEAKER_REFERENCE,
"language": LANGUAGE,
},
],
languages=[LANGUAGE],
)
# init the model from config
model = GPTTrainer.init_from_config(config)
# load training samples
train_samples, eval_samples = load_tts_samples(
DATASETS_CONFIG_LIST,
eval_split=True,
eval_split_max_size=config.eval_split_max_size,
eval_split_size=config.eval_split_size,
)
# init the trainer and 🚀
trainer = Trainer(
TrainerArgs(
restore_path=None, # xtts checkpoint is restored via xtts_checkpoint key so no need of restore it using Trainer restore_path parameter
skip_train_epoch=False,
start_with_eval=START_WITH_EVAL,
grad_accum_steps=GRAD_ACUMM_STEPS,
),
config,
output_path=OUT_PATH,
model=model,
train_samples=train_samples,
eval_samples=eval_samples,
)
trainer.fit()
if __name__ == "__main__":
main()
Inference Script
import os
import torch
import torchaudio
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts
# Get the directory of the current file
CURRENT_DIR = os.path.dirname(os.path.abspath(__file__))
TRAINING_DIR = os.path.join(CURRENT_DIR, "run/training")
TRAIN_CHECKPOINT_DIR = os.path.join(TRAINING_DIR, "GPT_XTTS_v2.0_LJSpeech_FT-April-03-2025_11+52AM-0000000")
# Add here the xtts_config path
CONFIG_PATH = os.path.join(TRAIN_CHECKPOINT_DIR, 'config.json')
# Add here the vocab file that you have used to train the model
TOKENIZER_PATH = os.path.join(CURRENT_DIR, 'tokenizer_ljs.json')
# Add here the checkpoint that you want to do inference with
XTTS_CHECKPOINT = os.path.join(TRAIN_CHECKPOINT_DIR, 'best_model.pth')
# Get the parent directory of the current file
PARENT_DIR = os.path.dirname(CURRENT_DIR)
SPEAKER_REFERENCE = os.path.join(PARENT_DIR, "LJSpeech-1.1/wavs/LJ001-0002.wav") # speaker reference to be used in training test sentences
# output wav path
OUTPUT_WAV_PATH = "xtts-ft_3.wav"
print("Loading model...")
config = XttsConfig()
config.load_json(CONFIG_PATH)
model = Xtts.init_from_config(config)
model.load_checkpoint(config, checkpoint_path=XTTS_CHECKPOINT, vocab_path=TOKENIZER_PATH, use_deepspeed=False)
print("Computing speaker latents...")
gpt_cond_latent, speaker_embedding = model.get_conditioning_latents(audio_path=[SPEAKER_REFERENCE])
print("Inference...")
out = model.inference(
"It took me quite a long time to develop a voice and now that I have it I am not going to be silent.",
"en",
gpt_cond_latent,
speaker_embedding,
temperature=0.7, # Add custom parameters here
)
torchaudio.save(OUTPUT_WAV_PATH, torch.tensor(out["wav"]).unsqueeze(0), 24000)
config.json file
{
"output_path": "/projects/coqui/LJTraining/run/training",
"logger_uri": null,
"run_name": "GPT_XTTS_v2.0_LJSpeech_FT",
"project_name": "XTTS_trainer",
"run_description": "\n GPT XTTS training\n ",
"print_step": 50,
"plot_step": 100,
"model_param_stats": false,
"wandb_entity": null,
"dashboard_logger": "tensorboard",
"save_on_interrupt": true,
"log_model_step": 1000,
"save_step": 10000,
"save_n_checkpoints": 0,
"save_checkpoints": false,
"save_all_best": false,
"save_best_after": 0,
"target_loss": null,
"print_eval": false,
"test_delay_epochs": 0,
"run_eval": true,
"run_eval_steps": null,
"distributed_backend": "nccl",
"distributed_url": "tcp://localhost:54321",
"mixed_precision": false,
"precision": "fp16",
"epochs": 1000,
"batch_size": 32,
"eval_batch_size": 16,
"grad_clip": 0.0,
"scheduler_after_epoch": true,
"lr": 5e-06,
"optimizer": "AdamW",
"optimizer_params": {
"betas": [
0.9,
0.96
],
"eps": 1e-08,
"weight_decay": 0.01
},
"lr_scheduler": "MultiStepLR",
"lr_scheduler_params": {
"milestones": [
900000,
2700000,
5400000
],
"gamma": 0.5,
"last_epoch": -1
},
"use_grad_scaler": false,
"allow_tf32": false,
"cudnn_enable": true,
"cudnn_deterministic": false,
"cudnn_benchmark": false,
"training_seed": 1,
"model": "xtts",
"num_loader_workers": 4,
"num_eval_loader_workers": 0,
"use_noise_augment": false,
"audio": {
"sample_rate": 22050,
"output_sample_rate": 24000,
"dvae_sample_rate": 22050
},
"use_phonemes": false,
"phonemizer": null,
"phoneme_language": null,
"compute_input_seq_cache": false,
"text_cleaner": null,
"enable_eos_bos_chars": false,
"test_sentences_file": "",
"phoneme_cache_path": null,
"characters": null,
"add_blank": false,
"batch_group_size": 48,
"loss_masking": null,
"min_audio_len": 1,
"max_audio_len": Infinity,
"min_text_len": 1,
"max_text_len": Infinity,
"compute_f0": false,
"compute_energy": false,
"compute_linear_spec": false,
"precompute_num_workers": 0,
"start_by_longest": false,
"shuffle": false,
"drop_last": false,
"datasets": [
{
"formatter": "",
"dataset_name": "",
"path": "",
"meta_file_train": "",
"ignored_speakers": null,
"language": "",
"phonemizer": "",
"meta_file_val": "",
"meta_file_attn_mask": ""
}
],
"test_sentences": [
{
"text": "It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.",
"speaker_wav": [
"/projects/coqui/LJTraining/../LJSpeech-1.1/wavs/LJ001-0002.wav"
],
"language": "en"
},
{
"text": "This cake is great. It's so delicious and moist.",
"speaker_wav": [
"/projects/coqui/LJTraining/../LJSpeech-1.1/wavs/LJ001-0002.wav"
],
"language": "en"
}
],
"eval_split_max_size": 256,
"eval_split_size": 0.01,
"use_speaker_weighted_sampler": false,
"speaker_weighted_sampler_alpha": 1.0,
"use_language_weighted_sampler": false,
"language_weighted_sampler_alpha": 1.0,
"use_length_weighted_sampler": false,
"length_weighted_sampler_alpha": 1.0,
"model_args": {
"gpt_batch_size": 1,
"enable_redaction": false,
"kv_cache": true,
"gpt_checkpoint": "",
"clvp_checkpoint": null,
"decoder_checkpoint": null,
"num_chars": 255,
"tokenizer_file": "/projects/coqui/LJTraining/tokenizer_ljs.json",
"gpt_max_audio_tokens": 605,
"gpt_max_text_tokens": 402,
"gpt_max_prompt_tokens": 70,
"gpt_layers": 30,
"gpt_n_model_channels": 1024,
"gpt_n_heads": 16,
"gpt_number_text_tokens": 21657,
"gpt_start_text_token": 0,
"gpt_stop_text_token": 1,
"gpt_num_audio_tokens": 1026,
"gpt_start_audio_token": 1024,
"gpt_stop_audio_token": 1025,
"gpt_code_stride_len": 1024,
"gpt_use_masking_gt_prompt_approach": true,
"gpt_use_perceiver_resampler": true,
"input_sample_rate": 22050,
"output_sample_rate": 24000,
"output_hop_length": 256,
"decoder_input_dim": 1024,
"d_vector_dim": 512,
"cond_d_vector_in_each_upsampling_layer": true,
"duration_const": 102400,
"min_conditioning_length": 66150,
"max_conditioning_length": 132300,
"gpt_loss_text_ce_weight": 0.01,
"gpt_loss_mel_ce_weight": 1.0,
"debug_loading_failures": false,
"max_wav_length": 255995,
"max_text_length": 200,
"mel_norm_file": "/projects/coqui/LJTraining/mel_mean_ljs.pth",
"dvae_checkpoint": "/projects/coqui/LJTraining/run/training/XTTS_v2.0_original_model_files/dvae.pth",
"xtts_checkpoint": "/projects/coqui/LJTraining/run/training/GPT_XTTS_v2.0_LJSpeech_FT-April-02-2025_09+37PM-0000000/model.pth",
"vocoder": ""
},
"model_dir": null,
"languages": [
"en"
],
"temperature": 0.85,
"length_penalty": 1.0,
"repetition_penalty": 2.0,
"top_k": 50,
"top_p": 0.85,
"num_gpt_outputs": 1,
"gpt_cond_len": 12,
"gpt_cond_chunk_len": 4,
"max_ref_len": 10,
"sound_norm_refs": false,
"optimizer_wd_only_on_weights": true,
"weighted_loss_attrs": {},
"weighted_loss_multipliers": {},
"github_branch": "inside_docker"
}
Expected behavior
No response
Logs
Environment
{
"CUDA": {
"GPU": [
"NVIDIA H100 80GB HBM3"
],
"available": true,
"version": "12.4"
},
"Packages": {
"PyTorch_debug": false,
"PyTorch_version": "2.7.0.dev20250226+cu124",
"TTS": "0.26.0",
"numpy": "1.26.4"
},
"System": {
"OS": "Linux",
"architecture": [
"64bit",
""
],
"processor": "x86_64",
"python": "3.11.11",
"version": "#1 SMP Thu Aug 8 17:47:08 UTC 2024"
}
}Additional context
No response