Access under PolyForm Small Business
Access is approved automatically after you confirm eligibility or written permission.
Read the PolyForm Small Business License 1.0.0 linked below before requesting access. Your Hugging Face username and email are shared with Trelis Research to manage access and send a welcome email.
Log in or Sign Up to review the conditions and access this model content.
Trelis Voice Isolator
Suppress background speech and noise in 16 kHz mono audio, running on CPU without a speaker-enrolment sample or cloud inference service.
Licence and access
Available under the PolyForm Small Business License 1.0.0, which permits commercial use by qualifying small businesses. Access is approved automatically after you confirm eligibility or written permission from Trelis Research.
If your company does not qualify, contact hello@trelis.com for a simple annual commercial licence.
Quick start
Requires x86-64 Linux with glibc, Python 3.11 or later, and uv.
Accept access on this page, authenticate and download:
uvx --from huggingface-hub==0.36.2 hf auth login
uvx --from huggingface-hub==0.36.2 hf download Trelis/voice-isolator --local-dir voice-isolator
uv run --no-project --with numpy==2.4.4 --with soundfile==0.13.1 python voice-isolator/enhance.py input.wav output.wav
Supply a 16 kHz mono WAV and a new output path. The example verifies the model and runtime hashes and writes a float32 WAV with the same sample count. It does not resample or compensate for model delay. For reproducible deployments, add --revision <full-commit-hash> to the download command.
Streaming use
Run from the downloaded directory with NumPy installed:
import numpy as np
from runtime import HushModel
model = HushModel("libweya_nc.so", "voice-isolator.tar.gz")
stream = model.stream()
try:
# Supply consecutive microphone chunks as mono float32 at 16 kHz.
microphone_chunk = np.zeros(320, dtype=np.float32)
enhanced_chunk = stream.process(microphone_chunk)
finally:
stream.close()
Keep one stream per microphone session and preserve its state across chunks. Incomplete 160-sample frames are buffered, so a call can return fewer samples than it receives. Reset only when starting an independent stream. Do not call the same stream concurrently from multiple threads.
For interruption detection, see the Barge-In Detector integration.
Benchmark
Comparison with upstream Hush using the same downstream interruption detector within each comparison:
| Measurement | Hush | Trelis Voice Isolator |
|---|---|---|
| Development false activations | 35 / 200 | 32 / 200 |
| Development interruptions detected | 84 / 160 | 118 / 160 |
| Evaluation false activations | 1 / 9 | 0 / 9 |
| Evaluation interruptions detected | 17 / 18 | 17 / 18 |
These internal results measure the combined enhancer and detector. The development set informed model selection; the evaluation set is small. They do not measure live latency.
Limitations
- Quiet or distant speech, short responses and speech onsets can be suppressed. Evaluate before using the model ahead of speech recognition.
- Nearby competing speakers, overlapping speech, television and reverberation can reduce separation quality. The model does not identify a particular speaker.
- Quality may vary across languages, accents and devices.
- The 10 ms processing frame size is not an end-to-end latency measurement.
Attribution and package
Derived from Hush. Upstream components retain their own licences; see third-party notices.
The model is a native runtime bundle loaded by runtime.py. File hashes are recorded in manifest.json.
Model tree for Trelis/voice-isolator
Base model
weya-ai/hush