| --- |
| title: VoxSplit |
| emoji: 🎙️ |
| colorFrom: purple |
| colorTo: pink |
| sdk: docker |
| app_port: 7860 |
| pinned: false |
| --- |
| |
| # Audio Transcription & Gender Detection (POC) |
|
|
| Upload a `.wav` file and get a **diarized, timestamped transcript** (via the |
| [Sarvam AI](https://sarvam.ai) `saaras:v3` batch API) plus a **per-speaker |
| gender estimate**. The transcript is synced to audio playback — the active |
| segment highlights as it plays, and clicking a segment seeks to it. |
|
|
| ## How it works |
|
|
| 1. The WAV is uploaded to a small FastAPI backend. |
| 2. The backend runs a Sarvam **batch STT job** with `with_diarization=True`, |
| which returns speaker-labelled segments with start/end timestamps. |
| 3. For each speaker, the backend pools all of their audio and runs the |
| [`prithivMLmods/Common-Voice-Gender-Detection`](https://huggingface.co/prithivMLmods/Common-Voice-Gender-Detection) |
| wav2vec2 classifier, returning softmax **female / male** probabilities. Below |
| a 0.6 confidence floor (or with too little audio) the speaker is marked |
| **uncertain**. |
| 4. The frontend renders the audio player, a speaker legend, and the synced |
| transcript. |
|
|
| > The gender model is downloaded from HuggingFace on first run (~360 MB) and |
| > cached. It's a trained classifier (~98% reported accuracy) but can still err |
| > on children, atypical voices, or noisy/short audio. |
|
|
| ## Setup |
|
|
| ```bash |
| cd audio-gender-detection |
| python3 -m venv .venv |
| source .venv/bin/activate |
| pip install -r requirements.txt |
| |
| cp .env.example .env # then add your Sarvam API key |
| ``` |
|
|
| `.env`: |
|
|
| ``` |
| SARVAM_API_KEY=your_sarvam_api_key_here |
| NUM_SPEAKERS=2 |
| ``` |
|
|
| (You can also paste the key directly into the UI instead of using `.env`.) |
|
|
| ## Run |
|
|
| ```bash |
| uvicorn backend.main:app --reload --port 8000 |
| ``` |
|
|
| Open http://localhost:8000 |
|
|
| ## Notes |
|
|
| - Diarization is **only** available through Sarvam's Batch API, so processing is |
| asynchronous — longer files take longer. |
| - Uploaded files land in `uploads/` (gitignored). Clean it up periodically. |
| - `librosa`/`soundfile` need a working audio backend; on macOS these install |
| cleanly via pip. On Linux you may need `libsndfile1` (`apt install libsndfile1`). |
|
|
