Inworld releases Voice Isolator v1: Apache-2.0 48 kHz voice isolation models at 64M and 58M parameters
Inworld has released Voice Isolator v1, two Apache-2.0 models for 48 kHz voice isolation: a 64M-parameter offline model and a 58M-parameter causal realtime model. The realtime model has 63 ms algorithmic latency at 20 ms of look-ahead. Both do denoising only; restoration is planned for version 2.
Inworld has released Voice Isolator v1, two Apache-2.0 models that isolate voice from noisy 48 kHz speech recordings and live audio. The offline model, voice-isolator-v1, has 64.0M parameters. The realtime model, voice-isolator-v1-realtime, has 58.2M parameters and an algorithmic latency of 63 ms at 20 ms of look-ahead. Code and weights are both Apache-2.0.
Two checkpoints, one architecture
Offline (voice-isolator-v1) is built for recorded audio and anything not latency-sensitive. It is two-sided: every frame sees its whole 8 s chunk, and long inputs are rendered in 8 s chunks overlapping by 2 s.
Realtime (voice-isolator-v1-realtime) is causal, with a trained look-ahead of 0, 20 or 40 ms. It is meant for live calls and streams.
Both models take one channel at 48 kHz. Both keep the foreground voice: when two people speak at once, the louder voice is kept and one a few dB quieter is removed, while voices at similar levels are usually both kept.
Architecture
The network is a band-split dual-path design over the STFT (n_fft 2048, hop 480 at 48 kHz, 10 ms frames). Each frame's spectrum is cut into 46 bands, with one token per band and frame at width 512. Seven blocks refine the tokens, alternating two kinds:
- Four blocks run time attention with rotary position and exclusive self attention, followed by attention across bands.
- Three blocks run Mamba along time, followed by Mamba across bands (state 16).
Per-band heads produce a bounded complex mask and a bounded per-bin residual, which is added to the input spectrum. The realtime model's blocks are causal and it carries a look-ahead embedding.
Training data
Inworld says v1 was trained on approximately 10,000 hours of unique audio. The set pairs recorded speech (read and conversational, many languages, expressive and character voices, whispered speech, singing, and non-verbal sounds such as laughter, breaths and sighs) with environmental and household noise, traffic, rain, wind, music, crowd sound, transient events and competing talkers. Mixes pass through measured and simulated room responses and telephony and codec degradations. Real recordings of the same speech, captured clean and in rooms, complete the set. Third-party datasets are credited in the NOTICE file.
Scope and limits
Version 1 does denoising (noise suppression). It removes what surrounds the voice and does not restore anything the recording lost, such as missing bandwidth. Restoration is planned for version 2. The models are not a speaker selector. Subtracting the isolated voice from the input yields the background noise or a backing-track residual.
Usage and hardware
Install with pip install inworld-voice-isolation. The voice-isolate CLI writes the isolated voice and the removed background as float32 WAVs at 48 kHz, with original levels preserved. Streaming mode is selected with --mode streaming --lookahead 2. Weights download on first use with checksum verification. The package uses CUDA when available, Apple's Metal backend (mps) on Apple-silicon Macs with PyTorch 2.6 or later, and the CPU otherwise.
Inworld tested on NVIDIA A100, H100, B200 and B300 GPUs and on CPU. Smaller GPUs are not tested yet. The interactive demo's online mode streams the microphone for up to 60 seconds.
Benchmark
Inworld also published VIBE (Voice Isolation Benchmark and Evals), which ranks 48 kHz voice isolators on a denoising board and a restoration board. Inworld scores every system on its internal evaluation set, which is not distributed.
What this means
The release gives builders an open-license option with separate checkpoints for offline and live use. The 63 ms algorithmic latency at 20 ms look-ahead is the figure to check for streaming pipelines. The denoising-only scope matters for anyone expecting bandwidth recovery, since Inworld defers that to version 2. Because VIBE scores come from an internal evaluation set that is not distributed, teams will need to test the models on their own audio.
Related Articles
Cloudflare releases open-weight Clairic 27B and new Clairic-flash 9B for data extraction
Cloudflare announced on November 18, 2026 an updated 27-billion-parameter Clairic and a new 9-billion-parameter Clairic-flash. Both have open weights on Hugging Face and are available as hosted Workers AI endpoints.
Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license
Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.
Bria releases RMBG-3.0, a 4-step background remover built on Fibo 2.0, with open weights for non-commercial use
Bria AI has released RMBG-3.0, the third generation of its background removal models and the first built on a generative foundation. It is Fibo 2.0, fine-tuned and distilled to produce an alpha matte in 4 steps. Bria claims it ranks first among six systems in its own benchmark of 234 hard images, with an average rank of 1.51.
KAIST MLP Lab releases KORMo-Web-11B: open LLM, VLM and web agent scoring 30.1 on Korean Online-Mind2Web
KAIST MLP Lab has released KORMo-Web-11B, an 11B-parameter bilingual model that combines text, vision and browser-agent capabilities in one set of weights under Apache-2.0. On 612 Online-Mind2Web tasks it scores 29.0 in English and 30.1 in Korean, according to the lab.
Comments
Loading...