EMTIAZZ/xtts-bangla-finetuned

XTTS v2 — Bangla Fine-tuned (Voice Cloning)

Fine-tuned Coqui XTTS v2 model for Bangla (Bengali) Text-to-Speech with voice cloning. Trained on authentic Bangladeshi speech data for 406,521 steps, enabling high-quality Bangla speech synthesis that preserves a reference speaker's voice characteristics.

Model Details

Property	Value
Base Model	Coqui XTTS v2
Architecture	GPT-based TTS + HiFi-GAN vocoder
Training Data	Custom Bangladeshi Bangla speech dataset
Training Steps	406,521
Training Platform	Modal (H100 GPU)
Capabilities	TTS + Voice Cloning
Language	Bangla (bn)

Key Features

• Voice Cloning: Synthesize speech in any target speaker's voice using a short reference audio clip (~6-12 seconds)
• Authentic Bangla Pronunciation: Trained specifically on Bangladeshi speech patterns and accent
• High-Quality Audio: HiFi-GAN vocoder produces natural-sounding output
• Production Ready: Deployed and tested in production at 406K+ training steps

Architecture

XTTS v2 uses:

1. GPT model — autoregressive token generation conditioned on text + speaker embedding
2. HiFi-GAN vocoder — converts mel-spectrograms to waveform audio
3. Speaker encoder — extracts voice characteristics from reference audio for cloning

Training Details

• Framework: Coqui TTS 0.22.0
• Dataset: Bangla speech recordings from a native Bangladeshi female speaker
• Hardware: H100 GPU on Modal cloud
• Checkpoint: best_model_406521.pth (5.3GB full model)
• Special config: Uses language="en" internally due to fine-tuning setup

Limitations

• Private model — not available for direct download
• Single-speaker primary training data (Bangla female voice)
• Requires ~6GB+ VRAM for inference

Related Models

Model	Type	Access
orpheus-3b-bangla-high-data-finetuning	Orpheus 3B LoRA	Public
orpheus-3b-bangla-small-data-finetuning	Orpheus 3B LoRA	Public
bangladeshi-bangla-tts-vits	VITS	Public

Author

Emtiaz Uddin Ahmed — GitHub · Portfolio