Haas-ja-v1.0 CoreML
A CoreML conversion of third-intelligence/Haas-ja-v1.0, a Japanese speech recognition model, for running on Apple Silicon with FluidAudio.
The files follow FluidAudio's Parakeet TDT layout.
| File | Precision | Notes |
|---|---|---|
Preprocessor.mlmodelc |
fp32 | 15 s of 16 kHz audio to 128 mel bins |
Encoder.mlmodelc |
fp16 | FastConformer, 24 layers |
Decoderv2.mlmodelc |
fp32 | one KDA prediction-network step |
Jointerv2.mlmodelc |
fp16 | TDT joint with argmax over tokens and durations |
vocab.json |
3072 SentencePiece pieces, blank id 3072 |
The prediction network is two Kimi Delta Attention layers instead of an LSTM. Its recurrent state and short-convolution caches are packed into FluidAudio's LSTM-shaped h/c tensors, each [64, 1, 640].
Usage
FluidAudio needs the haasJa model version, which is on the haas-ja branch of wkentaro/FluidAudio.
fluidaudiocli transcribe audio.wav --model-version haas-ja
Validation
On four test clips, greedy decoding of these models produced the same transcripts as NeMo running the original checkpoint in fp32. The original model runs its KDA layers with fla Triton kernels, which need CUDA. This conversion and the fp32 reference both use a PyTorch port of the KDA step instead. Accuracy has not been measured on a benchmark.
License
Apache-2.0, same as the original model.
Model tree for wkentaro/Haas-ja-v1.0-coreml
Base model
third-intelligence/Haas-ja-v1.0