Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
44 lines
2 KiB
Markdown
44 lines
2 KiB
Markdown
# Haas-ja-v1.0
|
||
|
||
[Haas-ja-v1.0](https://huggingface.co/third-intelligence/Haas-ja-v1.0) is a
|
||
Japanese ASR model. It reuses Parakeet-TDT's FastConformer encoder and TDT joint,
|
||
replacing the LSTM prediction network with two Kimi Delta Attention blocks.
|
||
Select the `haas_asr` family. Its registration reuses the Parakeet runtime and
|
||
selects the KDA predictor; checkpoint variant labels do not select behavior.
|
||
|
||
```bash
|
||
./build/debug/bin/audiocpp_cli --task asr --family haas_asr \
|
||
--model models/Haas-ja-v1.0-GGUF/haas-ja-v1.0-q8_0.gguf \
|
||
--backend cuda --session-option haas_asr.matmul_weight_type=q8_0 \
|
||
--audio japanese.wav --text-out transcript.txt --log
|
||
```
|
||
|
||
F16 and Q8_0 packages are available. The checkpoint is Japanese-only. The
|
||
encoder uses full-context attention; the family's buffered streaming interface
|
||
does not make it a native cache-aware streaming encoder. Long-form splitting
|
||
uses the existing Parakeet runtime policy rather than upstream's quiet-point
|
||
splitting policy.
|
||
|
||
To convert the official checkpoint, download `Haas-ja-v1.0.nemo` from the
|
||
upstream repository, then run:
|
||
|
||
```bash
|
||
python tools/community_models/convert_haas.py \
|
||
--source Haas-ja-v1.0.nemo --staging haas-converted \
|
||
--converter build/debug/bin/audiocpp_gguf --type q8_0 \
|
||
--output haas-ja-v1.0-q8_0.gguf
|
||
```
|
||
|
||
Conversion requires PyTorch, PyYAML, NumPy, safetensors, SentencePiece,
|
||
tokenizers, protobuf and Transformers (tested with Transformers 5.12.1 and
|
||
protobuf 5.29.5). NeMo and flash-linear-attention are needed only
|
||
to run the official Python implementation, not for conversion or C++ inference.
|
||
|
||
The existing `haas_asr.matmul_weight_type=q8_0` session option also keeps
|
||
load-time packed and scaled projections in Q8_0. With `native`, projections
|
||
constructed from float32 values retain float32 storage. Omit the override when
|
||
using the F16 package if that precision is preferred.
|
||
|
||
Timestamp word boundaries follow the tokenizer’s word markers. Continuous
|
||
Japanese text can therefore occupy one phrase-sized span; this does not add
|
||
Japanese morphological word segmentation.
|