Third Intelligence's Japanese speech recognition model: It follows the structure of NVIDIA Parakeet-TDT (FastConformer encoder + TDT decoder, which can skip several frames at once during decoding), replacing the LSTM in the prediction network with Kimi Linear's Kimi Delta Attention. Officially, it claims higher accuracy with almost no change in speed. Training used approximately 46,000 hours of public corpora (ReazonSpeech, Common Voice, YODAS) plus internal data. Weights were uploaded on 10 22, and it has not been included in the last 7 periods.
Upvotes 54 · Likes 55 · Downloads 1532
Reasons
- Officially tested on the self-built and publicly available asr-ja-bench v1.1 (5 public datasets, 23.38 hours, with evaluation data completely excluded from training), it achieved a weighted strict CER of 6.12%, the lowest among the open-source models tested, comparable to GPT-Transcribe (6.07%); MAI-Transcribe 2, Scribe v2, and Gemini 3.1 Pro are more accurate.
- Speed: Officially claimed to achieve RTFx 3,236 on a single H200, roughly as fast as Parakeet (3,407) but with a strict CER about 1.1 points lower, and approximately 15–17 times faster than Nemotron 3.5 Transcribe and Whisper large-v3; estimated based on on-demand pricing for H200 in the Japan region, transcribing one hour of audio costs approximately 0.003–0.004 USD.
- Comprehensive practical features: Batch transcription of long audio by segmenting at silence; custom vocabularies can be added for names and terms (spelling corresponds to pronunciation), officially claiming it takes about 1.3 times longer; paired with NVIDIA Nemotron-3-Diarization, it can produce transcripts with speaker labels.
Usage
- pip install "haas-asr @ git+https://github.com/third-intelligence/Haas-ASR.git", then model = haas_asr.load_model("third-intelligence/Haas-ja-v1.0"), audio = haas_asr.load_audio("example.wav"), model.transcribe([audio]); for recordings longer than 30 seconds, use haas_asr.transcribe_longform(model, audio, batch_size=64).
- Custom vocabulary: haas_asr.Vocabulary({"spelling": ["pronunciation (kana)"]}), passed as the vocabulary parameter to transcribe or transcribe_longform.
- To preview the results, you can use the long audio page of the official Space at third-intelligence-haas-ja-v1-0.hf.space/longform.
Risks
- Recognizes Japanese only; requires Linux + NVIDIA GPU + NeMo 2.7, the prediction network depends on Triton / CUDA kernels from flash-linear-attention, and it cannot run on CPU or macOS.
- Trained for offline full-sentence transcription, low-latency streaming requires further tuning; output style cannot be controlled, tends to omit modal particles, and the writing style for kanji, kana, and numbers follows the training data.
- Benchmarks are compiled by the official team (sample metadata and reconstruction scripts are publicly available); costs are estimated based on full GPU load, excluding operational overhead; speeds for commercial APIs are measured via OpenRouter, including network time, and are for reference only.