Third Intelligence's Haas-ja-v1.0 Beats GPT-Transcribe on Japanese Speech at 75x Lower Cost
A Japanese ASR model from Third Intelligence hits 6.12% CER at roughly 1/70th the cost of GPT-Transcribe, matching commercial APIs.
- Haas-ja-v1.0 hits 6.12% CER strict on Japanese ASR, best among open models evaluated.
- Costs roughly $0.003 per audio hour, about 1/70th of GPT-Transcribe at similar accuracy.
- Based on NVIDIA Parakeet-TDT with Kimi Delta Attention replacing the LSTM prediction network.
- 615.8M parameters, FastConformer encoder, Apache 2.0 license, trained on ~46,000 hours.
- 15-17x faster than Whisper large-v3 and Nemotron 3.5 Transcribe at better accuracy.
- Linux + NVIDIA GPU only; offline transcription with longform batching and custom vocabulary support.
Tokyo-based Third Intelligence has released Haas-ja-v1.0, a 615.8-million-parameter Japanese automatic speech recognition model designed for fast, self-hosted transcription. The weights and inference code are available under the Apache 2.0 License.
On the company’s five-dataset benchmark, Haas-ja-v1.0 achieved a strict character error rate of 6.12%, the best result among the evaluated open models. Its reported throughput of 3,236 times real time places it on the speed-accuracy Pareto frontier, meaning no evaluated alternative was both faster and more accurate.
6.12% CER, with cost caveats
The asr-ja-bench v1.1 evaluation contains 18,318 utterances and 23.38 hours of Japanese speech. Character error rate, or CER, measures insertions, deletions and substitutions at the character level, with lower scores indicating more accurate transcripts. Third Intelligence built jacer to report three versions of the metric:
- Strict CER: Scores the transcript exactly as written, including differences between kanji and kana.
- Flex CER: Ignores valid kanji-versus-kana variations.
- Kana CER: Converts text to readings before comparison, emphasizing whether the speech sounds were recognized correctly.
| System | Availability | Strict CER | Approximate cost per audio hour |
|---|---|---|---|
| Haas-ja-v1.0 | Open weights and code | 6.12% | $0.003 to $0.004 |
| GPT-Transcribe | Commercial API | 6.07% | About $0.30 |
| MAI-Transcribe 2 | Commercial API | <6.12% | About $0.11 |
| Scribe v2 | Commercial API | <6.12% | About $0.11 |
| Gemini 3.1 Pro | Commercial API | <6.12% | Not stated |
The published Haas estimate makes an hour of transcription roughly 75 to 100 times cheaper than GPT-Transcribe and 28 to 37 times cheaper than the cited MAI and Scribe prices. That estimate represents infrastructure cost under benchmark-style utilization, whereas API prices include hosting and service overhead. Idle GPU capacity, storage, engineering and operations will raise the effective cost for lightly used deployments.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
