Haas-ja-v1.0

Haas-ja-v1.0

Turns Japanese speech into text with a 615.8M speech recognition model built for speed and accuracy. For developers and teams who transcribe Japanese audio on a Linux machine with an NVIDIA GPU.

Made by third-intelligence 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: Haas-ja-v1.0 turns spoken Japanese into text, with speed and accuracy as its two goals. It builds on NVIDIA's Parakeet-TDT design and swaps in Kimi Delta Attention in its prediction network. The card reports a strict CER of 6.12% on a benchmark made from five public Japanese datasets.

Who it suits: It suits developers and teams who need Japanese audio turned into text at speed.

Industry | Who uses it | How they use it | Example
Media | Video teams | Transcribe Japanese interviews and clips | A script from a recorded interview
Business | Operations teams | Turn Japanese meetings into text | Notes from a recorded call
Research | Speech researchers | Compare Japanese recognizers | A test on the card's five-dataset benchmark

People who work in other languages, or on a Mac or a machine without an NVIDIA GPU, can skip it.

Where to use it:

• An NVIDIA GPU on a Linux box, with Python 3.11 or newer and NeMo 2.7, as the card requires
• A Hugging Face Space demo, which the card links
• Windows, macOS or Linux for downloading the files, though the card says CPU and macOS runs are not supported

Use case | Who | What they get back
Short clips | Developers | The Japanese text of a recording
Long recordings | Operations teams | Text from audio over 30 seconds, split at quiet points
Names and terms | Editors | A custom vocabulary so rare names are written as you want

How to set it up: You need these first:

• Linux with an NVIDIA GPU
• Python 3.11 or newer and NeMo 2.7
• Triton and CUDA, which the attention kernels depend on
• The haas-asr inference package, which the card installs from its code repository

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Install the inference package into your environment.

pip install -U huggingface_hub
hf download third-intelligence/Haas-ja-v1.0

How to use it: Load the model, load an audio file, and call transcribe. For recordings longer than 30 seconds use the long-form call, which splits the audio at quiet points and works through the pieces in batches.

import haas_asr

model = haas_asr.load_model("third-intelligence/Haas-ja-v1.0")
audio = haas_asr.load_audio("example.wav")
output = model.transcribe([audio])
print(output[0].text)

audio = haas_asr.load_audio("example.flac")
text = haas_asr.transcribe_longform(model, audio, batch_size=64)

Tool | What it does
haas_asr.load_model | Returns the model as a standard NeMo object, set to bfloat16 on the GPU
haas_asr.load_audio | Reads WAV, FLAC, MP3 or OGG, makes it mono and resamples it to 16 kHz
model.transcribe | Transcribes a list of audio clips
haas_asr.transcribe_longform | Splits long audio at quiet points and transcribes it in batches

Example: You load a recording of a meeting in Japanese and call transcribe on it. The model returns the spoken Japanese as text, and printing the first result shows it. A recording over 30 seconds goes through the long-form call instead.

For newcomers and experts: Speech recognition turns spoken words into written text. A first small thing to try is one short Japanese WAV file and the four lines of the first snippet.

For experts, the stack pairs a FastConformer encoder with a TDT decoder, as in Parakeet-TDT. The decoder also predicts how long each token lasts, so it can skip several encoder frames at a time. Kimi Delta Attention replaces the LSTM in the prediction network, which the card says raises accuracy at almost the same speed. The cost is a hard dependency on Triton and CUDA.

Load a Japanese audio file
It is made mono and resampled to 16 kHz
The encoder reads the audio frames
The decoder predicts tokens and durations
You get the Japanese text

Good to know:

Title: Parameters by part, in millions
Encoder: 608.9 M
Prediction network: 3.9 M
Joint network: 3.0 M

• Size: 615.8M parameters, with a tokenizer of 3,072 tokens
• Language: Japanese
• Variants: Hugging Face lists 14 quantized and 2 fine-tuned versions made by other people
• License: apache-2.0
• Cost to run: none named, no API key or account. The card says it costs a small fraction of commercial speech-to-text APIs
• Data it reaches: it downloads model files from Hugging Face and works on the audio files you give it
• Not for: CPU or macOS runs, which the card says are not supported, or languages other than Japanese

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: nemo
• Downloads in the last 30 days: 3,665
• Likes: 77
• License: apache-2.0
• Languages: ja
• Quantized versions: 14
• Fine-tunes: 2
• Tags: nemo, speech, audio, automatic-speech-recognition, japanese, fastconformer, conformer

Links

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

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download third-intelligence/Haas-ja-v1.0

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