Bhav-TTS: Emotional Bollywood Text-to-Speech

Bhav-TTS is a custom fine-tuned text-to-speech model built on top of F5-TTS (Flow Matching architecture), specialized for emotional Indian speech synthesis with Bollywood voice characteristics.

Model Details

• Base Model: F5-TTS (SWivid/F5-TTS)
• Fine-tuning Dataset: Bollywood emotional audio dataset
• Training Steps: 1,200,000+
• Model Size: 5.1 GB (model_last.pt)
• Sample Rate: 24 kHz
• Languages: Hindi, English (Indian English)
• Specialization: Emotional speech with Bollywood voice profiles

Features

✨ Emotion-Aware Synthesis

• Support for emotion tags in text (<happy>, <sad>, <angry>, etc.)
• Trained on emotionally-diverse Bollywood audio

🎭 Celebrity Voice Profiles

• Shahrukh Khan (Drama)
• Amitabh Bachchan (Baritone)
• Narendra Modi (Speech)
• Zero-shot voice cloning capabilities

🗣️ Multilingual Support

• Native Hindi speech synthesis
• Indian English accent
• Mixed-language capability

⚡ Production Ready

• Checkpoint trained to convergence
• Integrated into Gradio demo
• OpenAI-compatible API support

Usage

Basic Inference

from transformers import AutoModel
import soundfile as sf

# Load model
model = AutoModel.from_pretrained("your-username/bhav-tts")

# Generate speech
text = "नमस्ते! मैं बहुत खुश हूँ। <happy>"
audio = model(text, voice="hi_male")

# Save
sf.write("output.wav", audio, 24000)

With Reference Speaker

# Zero-shot voice cloning
ref_audio = "speaker_reference.wav"
ref_text = "कुछ संदर्भ टेक्स्ट"
text = "आपका जवाब यहाँ"

audio = model(text, ref_audio=ref_audio, ref_text=ref_text)

Training Configuration

Parameter	Value
Base Architecture	F5-TTS (ConvNeXt V2 + Transformer)
Dataset	Custom Bollywood emotional audio
Sample Count	1,000+ utterances
Training Steps	1,200,000+
Checkpoints	Multi-stage saving enabled

Model Files

• model_last.pt (5.1 GB) - Final trained model weights
• pretrained_model_1200000.pt (1.3 GB) - Intermediate checkpoint (1.2M steps)
• vocab.txt (13.8 KB) - Custom vocabulary/phoneme mapping

Citation

If you use Bhav-TTS in your research, please cite:

@article{chen2024f5tts,
  title={F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching},
  author={Chen, Yushen and others},
  journal={arXiv preprint arXiv:2410.06885},
  year={2024}
}

License

MIT License

Related Projects

• F5-TTS: https://github.com/SWivid/F5-TTS
• Svara-TTS: Multilingual Indic TTS
• IndicF5: AI4Bharat's native Indic language variant

---

Status: ✅ Production Ready | Date: June 30, 2026