---
library_name: audio.cpp
license: apache-2.0
language:
- ja
tags:
- audio.cpp
- gguf
- automatic-speech-recognition
base_model:
- third-intelligence/Haas-ja-v1.0
---

# Haas-ja-v1.0 GGUF

Follow [audio.cpp](https://github.com/0xShug0/audio.cpp) for runtime support.
Converted from [third-intelligence/Haas-ja-v1.0](https://huggingface.co/third-intelligence/Haas-ja-v1.0/tree/660728d39019e5e065046154beca44ac9a66a540).

Japanese FastConformer-TDT ASR with a Kimi Delta Attention predictor. Use the
`haas_asr` family with the F16 or Q8_0 GGUF.

```bash
audiocpp_cli --task asr --family haas_asr \
  --model 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
```

The encoder is full-context. Buffered streaming uses audio.cpp's Parakeet
interface; this is not a native cache-aware streaming checkpoint.

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.

## CUDA performance

RTX 5090, Debug audio.cpp build, one session with a short warmup followed by
24 Japanese JSUT test utterances (97.91 seconds total), one utterance per
request. audio.cpp uses Q8_0 weights and
`haas_asr.matmul_weight_type=q8_0`; official Haas Python uses its default
BF16 precision. Warmup is excluded from RTF. Peak VRAM is the observed process
peak across model loading and inference, sampled every 10 ms.

| Runtime | RTF ↓ | Peak VRAM |
| --- | ---: | ---: |
| Official Python | 0.00690 | 4,326 MiB |
| audio.cpp Q8_0 | 0.00379 | 1,335 MiB |

This workload was **1.82× faster** with **69% less peak VRAM** in audio.cpp.
These are a small workload's measurements, not a full-corpus benchmark.