Inworld Voice Isolator v1 β€” GGUF for CrispASR

GGUF conversions of inworld/voice-isolator (Apache-2.0, weights and code by Inworld AI) for the native ggml runtime in CrispASR (voice-isolator backend). Source revision 9b0285b3fe8403aa8e2f1dbb2165a80a4294c970; inference code inworld-ai/voice-isolator@c73b3fab.

file checkpoint weights voice SI-SDR vs upstream F32 output
voice-isolator-f16.gguf offline (denoiser.safetensors) F16 52.4 dB
voice-isolator-q8_0.gguf offline Q8_0 backbone 44.3 dB
voice-isolator-q4_k.gguf offline Q4_K backbone 26.3 dB
voice-isolator-realtime-f16.gguf realtime (causal, look-ahead 0/2/4) F16 56.2 dB (3 s causal forward)
voice-isolator-realtime-q8_0.gguf realtime Q8_0 backbone 46.7 dB

Per-band encoders and decoder heads stay at source precision in the quantized files.

crispasr --separate -m voice-isolator-q8_0.gguf -f noisy.wav   # -> noisy_voice.wav + noisy_background.wav (48 kHz mono)
crispasr --server -m asr.gguf --separate-model voice-isolator-q8_0.gguf   # POST /v1/audio/separation

Convert yourself: python models/convert-voice-isolator-to-gguf.py --input denoiser.safetensors --output voice-isolator-f16.gguf. See CrispASR docs/voice-isolator.md for the validation record. Third-party dataset notices: upstream NOTICE.

Downloads last month
343
GGUF
Model size
64M params
Architecture
voice-isolator
Hardware compatibility
Log In to add your hardware

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for cstr/inworld-voice-isolator-GGUF

Quantized
(3)
this model