1.1.2

• A voice sample of any length from 3 seconds up works: a longer sample gives its loudest 3 seconds in the first 30. stream(), isolate(sample=) and the Pipecat filter's voice_sample= used to fail on anything but exactly 3 seconds.
• isolate() returns a call shorter than 3 seconds unchanged, with a warning, instead of failing. A sample under 3 seconds, or live(sec=) under 3, is refused with a clear error.

1.1.1: live() locks on continuous speech

• live() and the Pipecat filter no longer stall when the caller talks without pausing.
• Digital silence at the start of a call no longer puts line noise into the voice sample.

VoxLock models for sherpa-onnx

VoxLock 1.1.1 ONNX models with meta data for sherpa-onnx's speech denoisers.

These are the ONNX files from huggingface.co/arbazsiddiqui/voxlock (revision a50f4e1d), with meta data added by scripts/voxlock/add_meta_data.py. The graphs and weights are unchanged: each model's graph is byte-identical to the original and its outputs are bit-identical.

enroll runs once on exactly 3 seconds of the caller's voice (the chunk model accepts only the enroll outputs of a 3-second sample). chunk runs on each chunk of the call and carries the LSTM state.

File	Role	Precision	Size	sha256
voxlock_enroll.onnx	enroll	fp32 (used with fp32, int8 and int8-full)	12.5 MB	5d06048c156bb8cd422d29848e6d38fb26a7b39f6f7fd92eccf2a6ed1ce003c6
voxlock_chunk_int8.onnx	chunk	int8, LSTMs in 8-bit (default)	66.6 MB	df61a9d0062b0adbcf1182302b1eb8b815f15ac47ea4923e8e164c1aff64121d
voxlock_chunk.onnx	chunk	fp32	87.4 MB	f02a548fcbdc921b8ff22760bc2e69a86eb098e70016e0d77a49f99defc25571
voxlock_chunk_int8_full.onnx	chunk	int8-full, LSTMs and matrix multiplies in 8-bit	57.7 MB	28648542825b778123f33fef91e76d27ec31f744651e6413d2297d47db53bc9e
voxlock_enroll_fp16.onnx	enroll	fp16 weights	6.3 MB	89a9949df98cb7ef7f133a19de828077dafbec4ab072d1f1d5f20134b904a79b
voxlock_chunk_fp16.onnx	chunk	fp16 weights	44.3 MB	ccbac121e74e2c4a52fdbff67e1af4b4594e1dd94ee1608ecfa3d0eb8e59c723
voxlock_test_call.wav	test input	16 kHz, 16-bit, 14 s	448 KB	16342ff0b5cefb64cc012a3185c641f8762a78803fa0d9f4966b288669b132b4
voxlock_test_voice_sample.wav	voice sample of the caller in the test input	16 kHz, 16-bit, 3 s	96 KB	a3cd4eba63f33ae1001a73144286e877802c141f8189c84ea1461b332817255f

Pairs: int8 = voxlock_enroll.onnx + voxlock_chunk_int8.onnx; int8-full = voxlock_enroll.onnx + voxlock_chunk_int8_full.onnx; fp16 = voxlock_enroll_fp16.onnx + voxlock_chunk_fp16.onnx; fp32 = voxlock_enroll.onnx + voxlock_chunk.onnx.

Meta data in both files: model_type=voxlock, model_role (enroll or chunk), version=1, model_version=1.1.1, precision, sample_rate=16000, n_fft=512, hop_length=128, window_length=512, window_type=hann, num_freq_bins=257, context_dim=512, voice_sample_seconds=3, model_url, hf_revision, code_url, license=Apache-2.0, comment. Chunk only: state_shape=6,32,256 (LSTM states h and c), frames_per_chunk=100, output_gain=-0.5.

VoxLock's weights are released under Apache-2.0.

The test call is a simulated room mixture of two LibriSpeech test-clean speakers (CC BY 4.0), from the VoxLock demo. The voice sample is the loudest 3 seconds of VoxLock's fp32 output on it.

1.1.0: 8-bit models, live(), Pipecat filter

• 8-bit models: VoxLock(precision=...) with int8 (the new default), int8-full, fp16 and fp32; voxlock --precision.
  int8 scores 5.11% word error rate on the Krisp benchmark against 4.95% for fp32, and streams in real time on fewer CPU threads.
• VoxLock.live(): a stream that takes the first 3 seconds of speech as the voice sample, passing the audio through until then.
• Pipecat audio filter, pip install "voxlock[pipecat]": voxlock.pipecat.VoxLockFilter. Tested with Pipecat 1.12.0.
• Output is at the input's level. 1.0.0 returned the caller at twice the level, so the command line could clip loud words.

VoxLock 1.0.0

The best performing voice isolator there is: 4.95% word error rate and 0.11 background words leaked per minute on the Krisp Voice Isolation Benchmark.

pip install voxlock
voxlock call.wav -o caller.wav

Model: https://huggingface.co/arbazsiddiqui/voxlock · Demo: https://arbazsiddiqui.me/voxlock