---
title: Audio Splitter with Diarization
emoji: 🎙️
colorFrom: blue
colorTo: blue
sdk: gradio
sdk_version: 6.10.0
python_version: "3.10.8"
app_file: app.py
full_width: true
pinned: false
license: mit
short_description: Separate each isolated voice for RVC training.
---

# Audio Splitter with Speaker Diarization

Split audio into segments organized by speaker. Uses faster-whisper for ASR with two speaker identification modes.

## Speaker ID Modes

### 1. Diarization (Default)
Best for **any number of speakers**. Labels who speaks when using clustering-based approach.

```
Audio → ASR (full audio) → Diarization → Assign speakers to segments
```

### 2. Speech Separation (2 speakers only)
Physically separates audio into **2 clean speaker tracks** using Dual-Path-RNN neural network.

```
Audio → Dual-Path-RNN → [Speaker1 track, Speaker2 track] → ASR on each
```

## Features

- **No HuggingFace token required** - Pure ONNX inference
- **Fast** - Optimized for CPU
- **Auto speaker detection** - Diarization mode detects speaker count automatically
- **Clean separation** - Speech separation mode provides isolated speaker tracks
- **Speaker Folders** - Output organized by speaker (SPEAKER_00/, SPEAKER_01/, etc.)
- **Paired Output** - Each segment has .wav + .txt files

## Output Structure

```
output/
├── SPEAKER_00/
│   ├── 0001.wav
│   ├── 0001.txt
│   └── ...
├── SPEAKER_01/
│   └── ...
└── transcript.txt
```

## Notes

- **Diarization mode** - Best for multi-speaker conversations with varying speaker counts
- **Separation mode** - Best for 2-speaker audio where clean tracks are needed (e.g., interviews)
- **faster-whisper** - Uses large-v3 model for best accuracy

## Usage

1. Upload audio file (MP3, WAV, etc.)
2. Select Whisper model (large-v3 recommended)
3. Enable speaker diarization
4. Set duration filter (Min/Max) - segments outside range are discarded
5. Click "Process"
6. Download ZIP with organized segments

### Duration Filter

- **Min (s)**: Discard segments shorter than this (filters noise/single words)
- **Max (s)**: Discard segments longer than this (useful for training data)

## Models Used

| Component | Model | Purpose |
|-----------|-------|---------|
| ASR | [faster-whisper](https://github.com/SYSTRAN/faster-whisper) | Speech-to-text with timestamps |
| Segmentation | [pyannote-segmentation-3.0](https://huggingface.co/onnx-community/pyannote-segmentation-3.0) | Detect speech regions |
| Embeddings | [3dspeaker](https://github.com/modelscope/3D-Speaker) | Speaker identity vectors |
| Clustering | sklearn AgglomerativeClustering | Group same speakers |
| Separation | [Dual-Path-RNN](https://github.com/JusperLee/Dual-Path-RNN-Pytorch) | Speech separation (2 speakers) |

## Credits

- [faster-whisper](https://github.com/SYSTRAN/faster-whisper) - CTranslate2 Whisper
- [pyannote-segmentation-3.0](https://huggingface.co/onnx-community/pyannote-segmentation-3.0) - ONNX segmentation
- [3D-Speaker](https://github.com/modelscope/3D-Speaker) - Speaker embeddings
- [Dual-Path-RNN](https://github.com/JusperLee/Dual-Path-RNN-Pytorch) - Speech separation
- [Original Audio Splitter](https://github.com/JarodMica/audiosplitter_whisper) by JarodMica