---
license: apache-2.0
library_name: pytorch
tags:
  - speech
  - asr
  - voice-activity-detection
  - turn-taking
  - speaker-verification
  - voice-agent
base_model:
  - nvidia/stt_en_fastconformer_hybrid_large_streaming_multi
  - nvidia/nemotron-speech-streaming-en-0.6b
---

# audioforge heads

Small trained heads for [audioforge](https://github.com/maxmelichov/audioforge), a real-time speech front end for
voice agents. They sit on top of one frozen NVIDIA streaming model and add voice activity, "is it the user?" and
"is the turn over?". These files are the heads only. The NVIDIA models are not redistributed here;
`audioforge-download` fetches them from NVIDIA under their own licences.

| file | core | what it is |
|---|---|---|
| `served_heads_v0.4.pt` | 115M (default) | VAD, turn, speaker and speech heads for `stt_en_fastconformer_hybrid_large_streaming_multi` |
| `served_heads_0p6b_v0.4.pt` | 0.6B | the same heads retrained on `nemotron-speech-streaming-en-0.6b` |
| `tsvad_spk.pt`, `tsvad_0p6b.pt` | 115M, 0.6B | target-speaker VAD head |
| `voice_gender_115m.pt`, `voice_gender_0p6b.pt` | 115M, 0.6B | perceived voice gender head |

## Use

```bash
git clone https://github.com/maxmelichov/audioforge && cd audioforge
uv run audioforge-download   # fetches the NVIDIA model and checks these heads
uv run audioforge-serve
```

The same files ship in the GitHub repo's `assets/`, so the download works without this page. Results and the test
protocol are in the repo's `docs/RESULTS.md`.

## Licence

The heads are Apache-2.0. The NVIDIA base models keep their own licences: CC-BY-4.0 for the 115M model (credit to
NVIDIA), NVIDIA Open Model License for the 0.6B model.