Whisper Large v3 Turbo - CoreML

Compiled CoreML bundle for OpenAI Whisper large-v3-turbo, staged for Apple-platform ASR in speech-swift.

Part of the soniqo.audio speech toolkit. This is the Apple CoreML bundle used by the native WhisperASR runtime in speech-swift. Browse Apple bundles in the CoreML Speech Models collection.

Model


Base model	openai/whisper-large-v3-turbo
Bundle source	argmaxinc CoreML bundle
Format	Compiled CoreML `.mlmodelc`
Runtime	speech-swift WhisperASR / CoreML
Total artifact size	1562.56 MiB
Conversion	Precompiled upstream CoreML bundle, staged for soniqo runtimes

The bundle splits Whisper into CoreML programs for mel extraction, audio encoding, context prefill, and token decoding. speech-swift loads these programs through its native WhisperASR runtime for Apple Silicon and iOS/macOS deployment.

Files

File	Size	Description
`AudioEncoder.mlmodelc/analytics/coremldata.bin`	0.0 MB	Audio encoder CoreML program
`AudioEncoder.mlmodelc/coremldata.bin`	0.0 MB	Audio encoder CoreML program
`AudioEncoder.mlmodelc/metadata.json`	0.0 MB	Audio encoder CoreML program
`AudioEncoder.mlmodelc/model.mil`	6.84 MB	Audio encoder CoreML program
`AudioEncoder.mlmodelc/model.mlmodel`	0.42 MB	Audio encoder CoreML program
`AudioEncoder.mlmodelc/weights/weight.bin`	1214.96 MB	Audio encoder CoreML program
`MelSpectrogram.mlmodelc/analytics/coremldata.bin`	0.0 MB	Mel spectrogram CoreML program
`MelSpectrogram.mlmodelc/coremldata.bin`	0.0 MB	Mel spectrogram CoreML program
`MelSpectrogram.mlmodelc/metadata.json`	0.0 MB	Mel spectrogram CoreML program
`MelSpectrogram.mlmodelc/model.mil`	0.01 MB	Mel spectrogram CoreML program
`MelSpectrogram.mlmodelc/weights/weight.bin`	0.36 MB	Mel spectrogram CoreML program
`TextDecoder.mlmodelc/analytics/coremldata.bin`	0.0 MB	Token decoder CoreML program
`TextDecoder.mlmodelc/coremldata.bin`	0.0 MB	Token decoder CoreML program
`TextDecoder.mlmodelc/metadata.json`	0.0 MB	Token decoder CoreML program
`TextDecoder.mlmodelc/model.mil`	0.13 MB	Token decoder CoreML program
`TextDecoder.mlmodelc/model.mlmodel`	0.11 MB	Token decoder CoreML program
`TextDecoder.mlmodelc/weights/weight.bin`	328.0 MB	Token decoder CoreML program
`TextDecoderContextPrefill.mlmodelc/analytics/coremldata.bin`	0.0 MB	Token decoder CoreML program
`TextDecoderContextPrefill.mlmodelc/coremldata.bin`	0.0 MB	Token decoder CoreML program
`TextDecoderContextPrefill.mlmodelc/metadata.json`	0.0 MB	Token decoder CoreML program
`TextDecoderContextPrefill.mlmodelc/model.mil`	0.0 MB	Token decoder CoreML program
`TextDecoderContextPrefill.mlmodelc/weights/weight.bin`	11.72 MB	Token decoder CoreML program
`config.json`	0.0 MB	Whisper / generation configuration
`generation_config.json`	0.0 MB	Whisper / generation configuration

Reference Benchmarks

M5 Pro, speech-swift native WhisperASR benchmark, full local FLEURS test split:

Dataset	WER	CER	Mean RTF	Overall throughput
`fleurs-en_us`	5.55%	2.72%	0.065	16.7x
`fleurs-fr_fr`	6.28%	2.70%	0.075	14.0x
`fleurs-ar_eg`	16.60%	5.63%	0.082	12.8x

Usage

Used by speech-swift through the native WhisperASR runtime:

speech transcribe /path/to/audio.wav \
  --engine whisper \
  --model aufklarer/Whisper-Large-v3-Turbo-CoreML \
  --language en