Whisper-Podlodka-Turbo - CoreML / WhisperKit
A CoreML conversion of
bond005/whisper-podlodka-turbo packaged for the
WhisperKit runtime. Runs end-to-end on the Apple Neural Engine (ANE) on Apple Silicon Macs, iPhone, and iPad.
The upstream model is Ivan Bondarenko's Russian-focused fine-tune of openai/whisper-large-v3-turbo, with improved noise robustness and reduced non-speech hallucinations. This repository contains only the converted weights - no architectural or training changes.
Used on-device in
Letopis, a local meeting and call recorder and transcriber for macOS and iOS.
Files
| File | Size | Purpose |
|---|
MelSpectrogram.mlmodelc | ~400 KB | Audio preprocessing (log-mel filterbank) |
AudioEncoder.mlmodelc | ~1.2 GB | 32-layer encoder, FP16 |
TextDecoder.mlmodelc | ~330 MB | 4-layer turbo decoder, FP16 |
config.json | - | Hugging Face Whisper config, inherited from the base model |
generation_config.json | - | Generation defaults, inherited from the base model |
All three .mlmodelc directories are pre-compiled MLProgram assets ready for direct use by WhisperKit. No additional compile step is required.
Architecture
Inherited from the base model (Whisper Large v3 Turbo):
| Hyperparameter | Value |
|---|
| Encoder layers | 32 |
| Decoder layers | 4 |
| Hidden size (d_model) | 1280 |
| Attention heads (enc / dec) | 20 / 20 |
| Mel bins | 128 |
| Vocabulary | 51866 |
| Max source positions | 1500 (30 s @ 16 kHz) |
| Max target positions | 448 |
Conversion
1whisperkit-generate-model \
2 --model-version bond005/whisper-podlodka-turbo \
3 --output-dir ./out \
4 --generate-decoder-context-prefill-data
- Conversion environment: Python 3.12,
torch==2.5.0, coremltools==9.0, transformers==4.53
- Compute precision: FP16 across all three components
- Decoder SDPA implementation:
Cat (default)
- Audio encoder SDPA implementation:
SplitHeadsQ (default)
- Decoder context prefill data: enabled (pre-computes the KV cache for the first 3 forced tokens to reduce time-to-first-token)
- Minimum deployment target: macOS 14 / iOS 17
Usage with WhisperKit (Swift)
1import WhisperKit
2
3let folder = URL(fileURLWithPath: "/path/to/whisper-podlodka-turbo-coreml")
4let pipe = try await WhisperKit(modelFolder: folder.path)
5let result = try await pipe.transcribe(audioPath: "/path/to/audio.wav")
6print(result?.text ?? "")
The tokenizer is the standard Whisper Large v3 tokenizer (vocab 51866) - WhisperKit will fetch it from openai/whisper-large-v3 if it is not present locally. Russian is the recommended decoding language for this model.
Performance
End-to-end ANE execution on an M-series Mac yields realtime factors significantly above 1.0x. First inference compiles ANE-specific kernels and may take noticeably longer; subsequent inferences use the cached compilation and are fast.
Languages
Primary: Russian. Secondary: English. The base fine-tune preserves the multilingual capability of Whisper Large v3 Turbo but is optimized for Russian ASR and Russian/English speech translation.
Limitations
- WhisperKit's pipeline currently uses the same scaffolding as standard Whisper Large v3 Turbo. Any quality differences between this fine-tune and the base Turbo model are inherited from the upstream weights.
- Translation behavior is inherited from the upstream fine-tune. Refer to the base model card for translation quality notes.
- For evaluation numbers (WER on Common Voice, RuLibriSpeech, Golos, SOVA RuDevices, Podlodka Speech, plus noise-robust and long-form benchmarks) see the upstream model card.
License
Apache 2.0, inherited from the base model.
Credits
Citation
For the base fine-tune, cite the upstream model:
1@misc{whisper-podlodka-turbo,
2 author = {Ivan Bondarenko},
3 title = {Whisper-Podlodka-Turbo: Enhanced Whisper Model for Russian ASR},
4 year = {2025},
5 publisher = {Hugging Face},
6 journal = {Hugging Face Model Hub},
7 howpublished = {\url{https://huggingface.co/bond005/whisper-podlodka-turbo}}
8}
