my-bangla-tts
A single-speaker Bengali (Bangla) text-to-speech model based on the VITS (Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech) architecture, fine-tuned from a base VITS checkpoint using the finetune-hf-vits toolkit.
This model converts Bengali text directly into natural-sounding speech waveforms in a single forward pass, without requiring a separate vocoder.
Model Details
- Architecture: VITS (
VitsModel, Transformers) - Language: Bengali (
bn) - Speakers: Single speaker (
num_speakers: 1) - Sampling rate: 16,000 Hz
- Vocabulary size: 74 tokens
- Framework: PyTorch, Transformers (
transformers==4.35.2) - Precision: float32
Architecture summary
| Component | Value |
|---|---|
| Hidden size | 192 |
| Flow size | 192 |
| Attention heads | 2 |
| Hidden layers | 6 |
| FFN dimension | 768 |
| Duration predictor | Stochastic |
| Prior encoder WaveNet layers | 4 |
| Posterior encoder WaveNet layers | 16 |
Usage
from transformers import VitsModel, VitsTokenizer
import torch
import scipy.io.wavfile as wavfile
model = VitsModel.from_pretrained("tareq052/my-bangla-tts")
tokenizer = VitsTokenizer.from_pretrained("tareq052/my-bangla-tts")
text = "আমার নাম তারেক, আমি একটি বাংলা ভয়েস মডেল তৈরি করেছি।"
inputs = tokenizer(text, return_tensors="pt")
with torch.no_grad():
output = model(**inputs).waveform
wavfile.write("output.wav", model.config.sampling_rate, output.squeeze().numpy())
Training
This model was fine-tuned on a Bengali speech dataset using the finetune-hf-vits toolkit, which adapts the original VITS architecture for fine-tuning within the Hugging Face transformers ecosystem.
Related Models
An emotion-conditioned variant of this model, fine-tuned on the BEASC (Bengali Emotional Speech Corpus) dataset with four emotions (angry, sad, happy, surprise) mapped to distinct speaker embeddings, is available separately.
Limitations
- Single-speaker only — this checkpoint does not support speaker or emotion conditioning.
- Trained on a limited dataset; pronunciation of rare words, numbers, or code-switched (non-Bengali) text may be inconsistent.
- Output sampling rate is fixed at 16 kHz.
License
Apache 2.0