---
language: bn
tags:
  - text-to-speech
  - bangla
  - vits
  - coqui-tts
license: apache-2.0
---

# Bangla VITS TTS

A Bengali (Bangla) text-to-speech model trained with [Coqui TTS](https://github.com/coqui-ai/TTS)
using the VITS architecture. This repository is updated automatically every **2 epochs**.

## Training progress

| Checkpoint | Loss | Step | Epoch |
|---|---|---|---|
| 🏆 Best model | `16.9650` | `26,910` | `630` |
| 📦 Current model | `17.5934` | `28,980` | `922` |

> Last updated: 2026-03-30 05:22 UTC

## Repository structure

| Path | Contents |
|---|---|
| `best_model/` | Checkpoint with the lowest validation loss seen so far |
| `current_model/` | Most recent checkpoint (updated every 2 epochs) |
| `training_state.json` | Latest step, epoch, and both losses — used to resume training |

## Quick start

```python
from TTS.api import TTS

# Best (lowest-loss) checkpoint
tts = TTS(
    model_path="best_model/<checkpoint>.pth",
    config_path="best_model/config.json",
)
tts.tts_to_file(
    text="আমি বাংলায় কথা বলতে পারি।",
    file_path="output.wav",
)
```
```python
from TTS.api import TTS

# Current (lowest-loss) checkpoint
tts = TTS(
    model_path="current_model/<checkpoint>.pth",
    config_path="current_model/config.json",
)
tts.tts_to_file(
    text="আমি বাংলায় কথা বলতে পারি।",
    file_path="output.wav",
)
```

## Training details

| Field | Value |
|---|---|
| Language | Bengali (`bn`) |
| Architecture | VITS |
| Framework | Coqui TTS |
| Sample rate | 22 050 Hz |

## Checkpoint policy

- **`current_model/`** is replaced on every push (every 2 epochs).
- **`best_model/`** is only replaced when the current loss is strictly lower than all
  previous checkpoints — it always contains the safest checkpoint for inference.

## License

Apache 2.0 — see [LICENSE](LICENSE) for details.