Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

inworld-ai/voice-isolator

Machine Learning

Open-license 48 kHz speech isolation models for offline recording cleanup and low-latency live streaming, with Python and CLI tools.

104
11
Python
Apache-2.0
4 days old

What it does

Voice Isolator is Inworld's open-license (Apache-2.0) speech-separation release. It ships two checkpoints built on the same architecture: voice-isolator-v1 (64M parameters) processes recordings offline, giving every frame view of its whole 8 s chunk; voice-isolator-v1-realtime (58M parameters) is a causal stream with 20 ms of look-ahead for live calls and streams. Both output float32 48 kHz WAV by default and keep the original levels, so the isolated voice plus the removed remainder reconstructs the aligned mono input. Subtracting the voice yields the background or backing-track residual, which can help with vocal/music separation.

It targets recording cleanup before playback or publication, speech-dataset curation and voice pipelines where speaker identity must survive. It is explicitly not a speaker selector or a restoration model; restoration is planned for version 2.

Quick start

Requires Python 3.10+ and PyTorch 2.1+. CUDA is used when available, Apple Metal (mps) on Apple-silicon Macs with PyTorch 2.6+, otherwise CPU.

pip install inworld-voice-isolation
# or from a clone: git clone https://github.com/inworld-ai/voice-isolator && cd voice-isolator && pip install -e .

curl -LO https://raw.githubusercontent.com/inworld-ai/voice-isolator/main/demo/examples/sample_14.wav

voice-isolate sample_14.wav sample_14_clean.wav

The first run downloads the offline model and checks it against the checksums in the model registry. Next to sample_14_clean.wav the CLI writes sample_14_clean_background.wav. Use --output-sr 0 to keep the input sample rate, --background <path> to choose where the removed audio goes, or --no-background to save only the voice.

Streaming can be done from Python by creating a session with lookahead_frames=2, pushing 960-sample (20 ms) blocks of 48 kHz mono audio and flushing at the end; the CLI equivalent is voice-isolate sample_14.wav sample_14_streamed.wav --mode streaming --lookahead 2. The interactive demo offers Offline and Online tabs, and the audio cleanup notebook walks through a first evaluation.

Constraints

  • Input and output are 48 kHz mono WAV by default; --device chooses CUDA, mps or CPU.
  • Streaming uses a fixed 20 ms look-ahead; the background can contain voice leakage, so it is not a clean noise-only track.
  • Hardware was tested on NVIDIA A100/H100/B200/B300 and CPU. Smaller GPUs and on-device targets beyond Apple silicon are untested and listed as future work in Plans.
  • The VIBE benchmark (voice isolation and restoration boards) is scored on an internal evaluation set that is not distributed; see how systems are scored.
  • Code and weights are Apache-2.0, with third-party notices in NOTICE.

HotRepo's take

This is a fresh, focused release: the repo appeared in HotRepo's new-project snapshot on November 20, 2026 with 104 stars and 11 forks, gaining 4 stars over the roughly 1.5-hour window recorded that day. No best daily rank is on record yet for this repo, and there is no Hacker News discussion attached. The technical draw is the pair of checkpoints sharing one architecture with a natural-language bandwidth: an offline model for best-quality cleanup and a realtime causal variant for live pipelines, both under a permissive license and shipped with CLI, Python API, demo and benchmark documentation.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. moonshine-ai/moonshineBuild low-latency, on-device voice interfaces with streaming speech-to-text, intent recognition, and text-to-speech.C++11,335
  2. n4ze3m/tiny-vocaA 4.5M-parameter English text-to-speech model with two voices that runs on CPU or in the browser.Python105
  3. microsoft/VibeVoiceA family of open-weight speech models for long-form transcription and expressive multi-speaker TTS.Python55,376
  4. pnnbao97/VieNeu-TTSVietnamese text-to-speech with instant voice cloning that runs on CPU or GPU for real-time speech.Python3,280
  5. OpenBMB/VoxCPMA 2B tokenizer-free TTS model for 30-language speech synthesis, voice design, and voice cloning.Python42,008