Silero VAD v6.2.1 CoreML

CoreML export of Silero VAD v6.2.1 for Apple platforms. The package includes a drop-in 32 ms streaming model and a faster 256 ms aggregate variant.

Model Table

Model	Format	Precision	Input	New audio per call	Size
`silero_vad.mlmodelc`	CoreML mlprogram	Float16	`[1, 1, 576]`	512 samples / 32 ms	640 KB
`silero_vad_256ms.mlmodelc`	CoreML mlprogram	Float16	`[1, 1, 4160]`	4096 samples / 256 ms	760 KB

Both variants use explicit LSTM state tensors:

Name	Shape	Description
`audio`	`[1, 1, 576]` or `[1, 1, 4160]`	64-sample left context plus current audio
`h`	`[1, 1, 128]`	LSTM hidden state
`c`	`[1, 1, 128]`	LSTM cell state
`probability`	`[1]`	Speech probability
`h_out`	`[1, 1, 128]`	Updated hidden state
`c_out`	`[1, 1, 128]`	Updated cell state

Files

File	Description
`silero_vad.mlmodelc/`	32 ms streaming CoreML bundle
`silero_vad_256ms.mlmodelc/`	256 ms aggregate CoreML bundle using noisy-OR over eight internal 32 ms windows
`config.json`	Runtime metadata, shapes, upstream version, and variant list

Performance

Benchmarked on an Apple Silicon Mac using a 50-file VAD set with 2,445.9 seconds of audio, onset 0.25, offset 0.20, min speech 0.25s, min silence 0.10s.

Model	File F1	Span F1	Precision	Recall	FAR	Miss Rate	Throughput
v5.1.2 CoreML baseline	100.00%	86.27%	100.00%	75.86%	0.00%	24.14%	410.0xRT
v6.2.1 `silero_vad.mlmodelc`	98.04%	87.47%	99.98%	77.74%	3.44%	22.26%	412.4xRT
v6.2.1 `silero_vad_256ms.mlmodelc`	92.59%	93.88%	99.97%	88.49%	7.80%	11.51%	1442.2xRT

The 32 ms variant preserves the existing streaming interface and is the safest drop-in replacement. The 256 ms variant is substantially faster and more recall-oriented, but uses coarser probability frames.

Usage

import CoreML

let config = MLModelConfiguration()
config.computeUnits = .all
let modelURL = Bundle.main.url(forResource: "silero_vad", withExtension: "mlmodelc")!
let model = try MLModel(contentsOf: modelURL, configuration: config)

For a complete Swift runtime, see:

Source

Converted from snakers4/silero-vad, tag v6.2.1.