Instructions to use starkdmi/VoiceIsolator_v1_Realtime_48K_MLX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use starkdmi/VoiceIsolator_v1_Realtime_48K_MLX with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download starkdmi/VoiceIsolator_v1_Realtime_48K_MLX --local-dir VoiceIsolator_v1_Realtime_48K_MLX
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
Voice Isolator v1 Realtime 48K (MLX)
This model was converted to MLX format from
inworld/voice-isolator
(denoiser_realtime.safetensors, revision 9b0285b3, released as
voice-isolator-v1-realtime).
Inworld's realtime voice isolator: 48 kHz mono, causal, with a trained look-ahead of 0, 20 or 40 ms (63 ms algorithmic latency at 20 ms), 58.2M parameters. It streams with a KV-cached causal attention and carried Mamba state.
Code: starkdmi/VoiceIsolatorMLX (MLX Swift). The offline model is starkdmi/VoiceIsolator_v1_48K_MLX.
| file | precision | size | |
|---|---|---|---|
model_fp16.safetensors |
fp16 | 139 MiB | recommended |
model_fp32.safetensors |
fp32 | 253 MiB | reference precision |
Measured on an Apple M4 (10-core GPU) with MLX Swift: a live 20 ms push takes 7.7 ms in fp16 and 9.9 ms in fp32 (median), with 20 ms of look-ahead. Against the PyTorch reference streaming session (CPU, float32) the fp32 output differs by at most 1.1e-6 across windowed rendering, live 20 ms pushes and look-ahead 0, 2 and 4 frames. The fp16 output is 65-75 dB SNR from fp32's on two demo clips.
Usage
import VoiceIsolatorMLX
let isolator = try VoiceIsolator(contentsOf: modelURL, mode: .streaming,
lookaheadFrames: 2) // fp16 compute by default
let session = try isolator.stream()
let ready = try session.push(samples) // 48 kHz mono, pieces of any size
let rest = try session.flush()
Format
Safetensors with the architecture and format in the metadata (config,
format: voice-isolator-mlx, format_version: 3). The 46 per-band encoders
and decoder heads are packed into padded batched tensors, the Mamba
directions are stacked and their output projections folded (in float64), and
the exact float32 bits of torch.hann_window(2048) are stored as
stft.window. The fp16 file keeps float32 for the scan parameters, the
Mamba output projections (in the offline model their pre-norm activations
exceed float16's range)
and the window.
Checksums
1a8d376e432319ff59be506b9f805287357d659d6a47d4cc51d46b68e0a4af11 model_fp16.safetensors
a72add45bcf0594e8bac631a62eca85d080677ac611612f26d2c898c58ee00a6 model_fp32.safetensors
License
Apache-2.0, as the original. The weights are Inworld AI's (Copyright 2026 Inworld AI); NOTICE carries their training-data attributions.
Quantized
Model tree for starkdmi/VoiceIsolator_v1_Realtime_48K_MLX
Base model
inworld/voice-isolator