Questions about training setup and model details

[manhcuong17072002]

Hello,

First of all, thank you so much for sharing this model!

I have a few questions regarding the training and usage details:

1. Training configuration – Could you please share some information about the training setup (GPU type, total training time, batch size, number of epochs, etc.)?
2. Audio format – What sampling rate was used for the training audio? (e.g., 16 kHz, 24 kHz, or 44.1 kHz?)
3. Inference performance – What is the current inference speed on your GPU? Have you tried using frameworks such as vLLM or others to optimize runtime performance?
4. Multi-speaker training – Have you experimented with multi-speaker training yet?
5. Dataset format – Would it be possible to release the dataset (or a subset) in a format like Parquet, containing fields such as audio | transcript | sampling_rate?

Thanks again for your contribution!

Best regards,

[pnnbao97]

Thank you so much for your interest in VieNeu-TTS and for the thoughtful questions! 😊

1. Training configuration
   I'll be sharing detailed training configuration (GPU type, batch size, epochs, etc.) in the GitHub repository soon. Stay tuned!
2. Audio format
   The training audio uses 16 kHz sampling rate. The model then upsamples to 24 kHz during inference for higher quality output.
3. Inference performance
   Current inference speeds vary depending on hardware. The model supports GGUF format (Q4 and Q8 quantized versions) which can significantly improve inference speed and reduce memory usage, especially on CPU. You can find these optimized formats in the model repository.
4. Multi-speaker training
   Yes! The model was trained with multi-speaker data, which is why it can perform voice cloning effectively with just a short reference audio sample.
5. Dataset format
   The dataset is already available at: https://huggingface.co/datasets/pnnbao-ump/VieNeu-TTS
   It contains the audio samples and transcripts used for training. Feel free to check it out!
   I'll be updating the GitHub repo with more detailed training configs and fine-tuning code soon. Thanks again for your support and great questions! 🙏