Haas-ja-v1.0-GGUF

Haas-ja-v1.0-GGUF

GGUF files of a Japanese speech recognition model that run only in speech.cpp, on macOS, Windows, Linux or the CPU alone. For people who transcribe Japanese audio on their own computer.

Made by sakasegawa on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds GGUF copies of Haas-ja-v1.0, a Japanese speech recognition model from Third Intelligence with 616M parameters. It is a FastConformer recognizer whose decoder uses Kimi Delta Attention layers. The files use speech.cpp's own layout, so they run only in speech.cpp 0.9.4 or later and not in llama.cpp, LM Studio or Ollama.

Who it suits: People who want to turn Japanese speech into text on their own machine, and developers who want a local recognizer with a command line and an HTTP server.

Industry | Who uses it | How they use it | Example
Meetings and notes | Teams with Japanese recordings | Transcribe audio locally, with voice detection for long files | A recorded meeting saved as a WAV file
Software development | Developers adding speech to text | Run the local server and call its HTTP API | A transcription endpoint on their own machine
Language research | People studying Japanese speech recognition | Try several file sizes on a CPU | Compare the smaller and larger files on one clip

You can skip it if you need another language or want to load the files in a tool other than speech.cpp.

Where to use it:

• speech.cpp 0.9.4 or later on macOS with Metal
• speech.cpp on Windows and Linux with Vulkan
• speech.cpp on the CPU alone, with no NVIDIA GPU needed
• A local server started with the speech serve command

Use case | Who | What they get back
Transcribe a short clip | Anyone with a WAV file | The recognized Japanese text
Transcribe a long recording | Note takers | Text for the regions where someone speaks
Recognize special words | People with names or jargon | Better odds that a custom word appears in your spelling

How to set it up: You need speech.cpp first, since nothing else reads these files. After that, the speech command fetches the model files from this repository the first time you run it.

• speech.cpp 0.9.4 or later, installed by its own instructions
• Enough disk space for the file type you choose

1. Install speech.cpp.
2. Run any command from the next section with the model name.
3. Let it fetch the Q8_0 file on that first run, or name another type.

If you want the repository files yourself, Hugging Face gives these steps:

pip install -U huggingface_hub
hf download sakasegawa/Haas-ja-v1.0-GGUF

How to use it: You run the speech command with the model name and an audio file. You can also give the path of a file you already downloaded. A type other than Q8_0 is picked by adding a colon and the type to the name, such as q4_k.

speech asr haas-ja-v1.0 utterance.wav
speech asr haas-ja-v1.0 --timestamps utterance.wav
speech asr haas-ja-v1.0 --vad silero-vad meeting.wav
speech asr haas-ja-v1.0 --keywords $'Haas=ハース\nAIモデル' utterance.wav
speech serve haas-ja-v1.0 --open

Tool | What it does
speech asr | Transcribes an audio file with the model
--timestamps | Adds times to the output
--vad silero-vad | Recognizes only the spoken regions, best for recordings over about 30 seconds
--keywords | Raises the odds of custom words, one per line, with optional readings after an equals sign
speech serve | Starts a local server with an HTTP API

Example: You run the recognizer on meeting.wav with voice detection switched on. The model returns the Japanese text for the stretches where someone speaks. Adding keywords makes it favor names you list, and that run returns the text without times.

For newcomers and experts: If you are new, install speech.cpp and run the first command on a short WAV file. Check the text that comes back before you try keywords or the server.

For experts, the model is a FastConformer encoder with a TDT decoder whose prediction network uses Kimi Delta Attention layers. It decodes greedily, as NeMo does by default. Keywords are favored while decoding with a weight of 2 unless you change it, and a keyword run decodes twice. Smaller file types save space and are generally less exact, though the card gives no accuracy numbers.

Audio file in
speech.cpp fetches the GGUF file
FastConformer encoder, TDT decoder
Japanese text out

Good to know:

Title: File size by type in GB
F16: 1.23 GB
Q8_0: 0.66 GB
Q6_K: 0.51 GB
Q5_K: 0.43 GB
Q4_K: 0.35 GB

• Runtime: only speech.cpp 0.9.4 or later reads these files
• Older version: speech.cpp 0.9.3 fetches an earlier revision without the custom vocabulary settings
• Language: Japanese only
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the audio files you give it, and a download of the model files on first use
• Not for: llama.cpp, LM Studio, Ollama or other programs that read GGUF

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Parameters: 616M
• Weights: about 308 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 5,149
• Likes: 7
• License: apache-2.0
• Languages: ja
• Based on: third-intelligence/Haas-ja-v1.0
• Tags: gguf, ggml, speech.cpp, fastconformer, automatic-speech-recognition

Links

• Model: https://huggingface.co/sakasegawa/Haas-ja-v1.0-GGUF
• Files: https://huggingface.co/sakasegawa/Haas-ja-v1.0-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sakasegawa/Haas-ja-v1.0-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.