Minimal setup to train F5-TTS via CLI with LJSpeech dataset

[Jamil-Palma]

Checks

• This template is only for usage issues encountered.
• I have thoroughly reviewed the project documentation but couldn't find information to solve my problem.
• I have searched for existing issues, including closed ones, and couldn't find a solution.
• I am using English to submit this issue to facilitate community communication.

Environment Details

Hi, thank you so much for this amazing project!

I’m trying to fine-tune F5-TTS using the LJSpeech dataset, and I followed all the instructions and preprocessing steps from prepare_ljspeech.py.
🙏 If possible, could someone provide a minimal and clean example of how to launch training directly via CLI using a custom dataset like LJSpeech?

Just the minimum config/command combination to make it work properly with the char tokenizer, without relying on the dynamic vocab path logic.

That would be very helpful for understanding the correct override structure or config needed for simple use cases like mine.

Thanks a lot in advance! 🙌

OS & Container
Docker Image: ghcr.io/swivid/f5-tts:main
Base OS: Ubuntu (inside Docker)

Python

Python version: 3.11.9

Dependencies (installed via Docker image)

PyTorch: 2.2.1 (CUDA 12.x)
Transformers, Datasets, Accelerate, etc. — as included in the Docker image

Hardware

GPU: NVIDIA RTX (via nvidia-docker, with CUDA support)
Mixed precision: enabled (fp16 via accelerate)

Dataset

Dataset: LJSpeech-1.1
Processed to: /workspace/F5-TTS/data/LJSpeech_char

Command used to launch training

accelerate launch --mixed_precision=fp16 src/f5_tts/train/train.py --config-name F5TTS_LJSpeech.yaml

Steps to Reproduce

1. Clone the repo (inside Docker container)

git clone https://github.com/SWivid/F5-TTS.git
cd F5-TTS

2. Download and extract LJSpeech dataset

wget https://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2
tar -xjf LJSpeech-1.1.tar.bz2

3. Preprocess dataset

python src/f5_tts/train/datasets/prepare_ljspeech.py

**Output:**

For LJSpeech_char, sample count: 13100
For LJSpeech_char, vocab size: 75
For LJSpeech_char, total 23.92 hours

```bash
# 4. Create custom config
cp src/f5_tts/configs/F5TTS_v1_Base.yaml src/f5_tts/configs/F5TTS_LJSpeech.yaml

Edited config (F5TTS_LJSpeech.yaml):

datasets:
  name: LJSpeech_char
  dataset_name: LJSpeech_char
  dataset_dir: /workspace/F5-TTS/data/LJSpeech_char
  tokenizer: char
  vocab_path: /workspace/F5-TTS/data/LJSpeech_char/vocab.txt
  batch_size_per_gpu: 38400
  batch_size_type: frame
  max_samples: 64
  num_workers: 16

model:
  name: F5TTS_v1_Base
  tokenizer: char

# 5. Launch training
accelerate launch --mixed_precision=fp16 src/f5_tts/train/train.py --config-name F5TTS_LJSpeech.yaml

Observed error:

FileNotFoundError: No such file or directory: '/workspace/F5-TTS/F5-TTS/src/f5_tts/../../data/LJSpeech_char_char/vocab.txt'

---

Let me know if you'd like me to upload sample audio/text files or additional logs.
Everything was done inside Docker using ghcr.io/swivid/f5-tts:main.

✔️ Expected Behavior

When using a custom dataset configuration (for example, using LJSpeech with the "char" tokenizer), the training script should:

Load the dataset from the specified dataset_dir (e.g., /workspace/F5-TTS/data/LJSpeech_char).

Use the provided vocab_path (e.g., /workspace/F5-TTS/data/LJSpeech_char/vocab.txt) for building the tokenizer.

Not dynamically generate a vocabulary path by concatenating the dataset name with the tokenizer type (e.g., it should not look for /workspace/F5-TTS/F5-TTS/src/f5_tts/../../data/LJSpeech_char_char/vocab.txt).

Successfully start the fine-tuning process without encountering a FileNotFoundError.

In summary, the expected behavior is that the model uses the exact configuration parameters specified in the custom YAML file, ensuring that the correct vocabulary file is loaded and training proceeds normally.

❌ Actual Behavior

root@f876255569f3:/workspace/F5-TTS/F5-TTS# accelerate launch --mixed_precision=fp16 src/f5_tts/train/train.py --config-name F5TTS_LJSpeech.yaml
Error executing job with overrides: []
Traceback (most recent call last):
File "/workspace/F5-TTS/F5-TTS/src/f5_tts/train/train.py", line 31, in main
vocab_char_map, vocab_size = get_tokenizer(tokenizer_path, tokenizer)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/F5-TTS/F5-TTS/src/f5_tts/model/utils.py", line 113, in get_tokenizer
with open(tokenizer_path, "r", encoding="utf-8") as f:
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: '/workspace/F5-TTS/F5-TTS/src/f5_tts/../../data/LJSpeech_char_char/vocab.txt'

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.
Traceback (most recent call last):
File "/opt/conda/bin/accelerate", line 8, in
sys.exit(main())
^^^^^^
File "/opt/conda/lib/python3.11/site-packages/accelerate/commands/accelerate_cli.py", line 48, in main
args.func(args)
File "/opt/conda/lib/python3.11/site-packages/accelerate/commands/launch.py", line 1194, in launch_command
simple_launcher(args)
File "/opt/conda/lib/python3.11/site-packages/accelerate/commands/launch.py", line 780, in simple_launcher
raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd)
subprocess.CalledProcessError: Command '['/opt/conda/bin/python', 'src/f5_tts/train/train.py', '--config-name', 'F5TTS_LJSpeech.yaml']' returned non-zero exit status 1.
root@f876255569f3:/workspace/F5-TTS/F5-TTS#

[SWivid]

oh you only use to set like

datasets:
  name: LJSpeech
  batch_size_per_gpu: 38400
  batch_size_type: frame
  max_samples: 64
  num_workers: 16

model:
  name: xxx  # model name
  tokenizer: char # tokenizer type
  tokenizer_path: null  # if tokenizer = 'custom', define the path to the tokenizer you want to use (should be vocab.txt)

just follow the provided example