README.md

---
title: VoiceForge Studio
emoji: 🎙️
colorFrom: indigo
colorTo: purple
sdk: gradio
app_port: 7860
pinned: false
---

# 🎙️ VoiceForge Studio

A professional, modern web application for Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) powered by the latest open-source models.

## Features

### 🗣️ Text-to-Speech
- **Voice Cloning** — Clone any voice from a reference audio sample
- **Voice Design** — Create unique voices from text descriptions
- **Emotion Control** — Adjust emotional tone (happy, sad, angry, etc.)
- **Multi-speaker** — Generate conversations with multiple speakers
- **Multiple Models**: Qwen3-TTS, VibeVoice, Voxtral, IndexTTS-2, OmniVoice

### 🎙️ Speech-to-Text (ASR)
- **Word-Level Timestamps** — Precise timing for every single word
- **Line-Level Timestamps** — Chunk-level timing for readability
- **Speaker Diarization** — Identify who spoke when (pyannote 3.1)
- **Speaker Identification** — Automatic speaker counting and labeling
- **Multiple Models**: Whisper Turbo, Qwen3-ASR, VibeVoice-ASR

### 📦 Batch Processing
- Process multiple audio files at once
- Export results as JSON
- Automatic diarization across all files

## Models Supported

| Type | Model | Size | Key Features |
|------|-------|------|-------------|
| TTS | Qwen3-TTS | 1.7B | Voice Design + Clone, 12 languages |
| TTS | VibeVoice 1.5B | 1.5B | Long-form 90min, 4 speakers |
| TTS | VibeVoice Realtime | 0.5B | Streaming, low latency |
| TTS | Voxtral 4B | 4B | vLLM, OpenAI-compatible |
| TTS | IndexTTS-2 | — | Emotion control (8 dims) |
| TTS | OmniVoice | — | OpenAI-compatible API |
| ASR | Whisper Turbo | — | Word timestamps, 99 langs |
| ASR | Qwen3-ASR 1.7B | 1.7B | Forced alignment, line stamps |
| ASR | Qwen3-ASR 0.6B | 0.6B | Fast, lightweight |
| ASR | VibeVoice-ASR | — | Native diarization + timestamps |
| Diarization | pyannote 3.1 | — | Speaker diarization |

## Usage

1. Select a TTS or ASR model from the dropdown
2. For TTS: Enter text, optionally upload a reference voice for cloning
3. For ASR: Upload audio, choose timestamp granularity and diarization options
4. Click Generate / Transcribe and view the beautiful output

## License

Models follow their respective licenses (MIT, Apache-2.0, etc.)