# Haas-ja-v1.0 Hits 6.12% CER on Japanese Speech [2026] > Source: > Published: 2026-11-03 10:16:31+00:00 Third Intelligence has released **Haas-ja-v1.0**, an open Japanese speech recognition model built for fast, accurate transcription. In its release post, researcher Tatsu Matsushima wrote, “We have developed a new Japanese speech recognition model, Haas-ja-v1.0, and released it as open source!”, as [GIGAZINE reported on November 2](https://gigazine.net/gsc_news/en/20261102-haas-ja-v1-asr). The dated announcement puts the release on October 30, 2026. The numbers are compact. The model has 615.8 million parameters and posts a 6.12% strict character error rate (CER) on a benchmark assembled from five public Japanese datasets. The model card says it is the most accurate open model its authors evaluated. Weights and inference code ship under the Apache License 2.0, so teams can download, run, and modify them without a license negotiation. This analysis covers what the release confirms, what it leaves open, and why a mid-sized Japanese-only model matters to anyone building voice products. It also places Haas-ja-v1.0 in the wider shift toward specialized open models, a trend we tracked in our [open-weight AI models spec sheet](https://shattered.io/open-weight-ai-models-side-by-side-the-spec-sheet/). A word on method: every figure below comes from the announcement coverage and the model documentation. Where the sources stay silent, this piece says so instead of filling the gap. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## What Third Intelligence Released on October 30 Third Intelligence is a Japanese AI company, and Matsushima is the researcher who signed the announcement. The product is a speech-to-text model that targets one language. That choice defines the whole release. Instead of covering dozens of languages with one set of weights, the team spent its parameter budget on Japanese alone. The release has three parts: the model weights, the inference code, and public demos. All of it is open. GIGAZINE’s coverage notes that the weights and code carry the Apache License 2.0, which is among the most permissive licenses in common use for machine learning artifacts. A company can embed the model in a commercial product, host it on its own servers, or fine-tune it on private audio, all without paying Third Intelligence. The official description frames the goal as a balance of two properties: speed and accuracy. Plenty of ASR (automatic speech recognition) systems pick one. Large models tend to win on accuracy and lose on latency and cost. Small models flip the trade. Haas-ja-v1.0 aims to sit in the middle, with a parameter count (615.8 million) that is modest by 2026 standards yet large enough to carry a serious acoustic encoder. What the release does not include matters too. The available sources establish no price, no subscription tier, and no hosted commercial API. Anyone who wants to use the model today runs it themselves or tries the demos Matsushima pointed to. Any cost claim you see elsewhere is not part of the confirmed record. ## Haas-ja-v1.0 Specs at a Glance The table below lists only what the announcement coverage and model documentation confirm. Items the sources do not establish are marked as such. | Attribute | Confirmed detail | |---|---| | Model name | Haas-ja-v1.0 | | Developer | Third Intelligence | | Release date | October 30, 2026 | | Parameters | 615.8 million | | Base design | NVIDIA Parakeet-TDT style | | Encoder | FastConformer | | Decoder | Token-and-Duration Transducer (TDT) | | Prediction network | Kimi Delta Attention (KDA) | | Benchmark result | 6.12% strict CER across five public Japanese datasets | | License | Apache License 2.0 (weights and inference code) | | Price or commercial plan | Not established by available sources | Two details stand out. First, the design borrows from NVIDIA’s Parakeet-TDT family but swaps in a different component in the prediction network. Second, the accuracy figure comes from a combined benchmark, not from one favorable test set. Both choices suggest the authors wanted a result that holds up under scrutiny. ## How the Parakeet-TDT Design Works Understanding why this model is fast requires a short tour of its parts. A transducer-style ASR system has three pieces: an encoder that turns audio into a sequence of acoustic features, a prediction network that tracks the text produced so far, and a joint component that combines both to pick the next output. Haas-ja-v1.0 follows that pattern with specific choices at each stage. ### The FastConformer encoder The encoder is a FastConformer. Conformer encoders blend convolution, which captures local sound patterns, with attention, which captures longer range context. The “fast” variant compresses the audio sequence early so the rest of the network processes fewer time steps. Fewer steps mean less compute per second of audio. For long recordings such as meetings or broadcasts, that saving compounds quickly. ### The TDT decoder and Kimi Delta Attention The decoder is a Token-and-Duration Transducer. A standard transducer emits one output per time step, including many blank steps. A TDT predicts the token and also how many audio frames that token spans, which lets the decoder skip ahead instead of walking through every frame. That is the main source of speed in this family of models. Third Intelligence then changed the prediction network to use Kimi Delta Attention. The sources confirm the choice but the fact sheet does not detail the reasoning, so this article will not guess at the authors’ motives. What the choice does show is that the team is borrowing ideas from recent language model research and applying them to speech. That cross-pollination is a pattern worth watching across 2026 releases. ## Reading the 6.12% Character Error Rate Character error rate counts the characters a system gets wrong relative to a reference transcript. The calculation adds substitutions, deletions, and insertions, then divides by the number of characters in the reference. A CER of 6.12% means roughly six characters in every hundred needed correction on the benchmark. Lower is better. CER is the right yardstick for Japanese. English ASR usually reports word error rate, but Japanese text has no spaces between words, so word boundaries depend on a tokenizer and the choice of tokenizer changes the score. Characters give a cleaner, tokenizer-free measure. The model card calls the figure a “strict” CER. The sources do not spell out the text normalization rules behind that label, so readers should avoid comparing the 6.12% against numbers from other papers without checking how each team normalized punctuation, numerals, and script variants. Japanese offers many ways to write the same sound, including hiragana, katakana, kanji, and Arabic or kanji numerals. A scoring script that treats those variants as errors will produce a higher CER than one that merges them. The five-dataset benchmark helps here. A single test set can flatter a model that happens to match its recording style, such as clean read speech. Pooling five public sets mixes conditions and reduces the odds that one lucky domain drives the headline. It does not eliminate the risk, because a pooled benchmark can still differ sharply from a company’s own call center audio or noisy field recordings. ## What “Most Accurate Open Model” Does and Does Not Mean The headline on the GIGAZINE piece says the model achieved the highest performance among open models. The confirmed wording is narrower. The model card states that Haas-ja-v1.0 was the most accurate open model evaluated by its authors, on their benchmark. That is a claim about a specific field of competitors, not a universal ranking. | Claim | Status | What backs it | |---|---|---| | Most accurate open model among those the authors evaluated | Confirmed as the authors’ statement | Model card and GIGAZINE coverage | | 6.12% strict CER on a five-dataset benchmark | Confirmed as reported | Model card and announcement | | Best open Japanese ASR model on every benchmark | Not established | No source makes this claim | | Open weights and code under Apache 2.0 | Confirmed | Announcement and coverage | | Any price or commercial plan | Not established | No source gives one | | Independent replication of the benchmark | Not yet reported | No third-party result in the sources | This distinction is routine in model releases and still worth stating. Authors choose the comparison set, the benchmark mix, and the scoring rules. Those choices are legitimate, and they are also the reason independent replication carries weight. Within days of an open release, community members usually run their own audio through the model and post results. Until those numbers appear, treat 6.12% as a credible first-party figure, not a settled fact about the field. The same caution applied to the wave of open-model launches we covered in [our look at the Qwen4 open-source promise](https://shattered.io/qwen4-launch-alibaba-open-source-promise-doubt-470-models-2026/). A bold headline and a careful model card often say different things, and the model card is the document to read. ## Why Japanese Is a Hard Language for Speech Recognition Japanese speech recognition punishes shortcuts. The language mixes three writing systems in ordinary sentences. It contains a large number of homophones, so one spoken sound can map to several written forms, and only context picks the right one. Speakers also drop subjects and particles in casual speech, which strips away clues a model would otherwise use. A multilingual model that spreads its capacity over many languages often pays a tax on any single one. Japanese gets only a slice of the training mix, and the vocabulary has to cover kanji alongside every other script. A Japanese-only model removes that competition. The tokenizer, the training data, and the evaluation can all focus on one target. That does not guarantee a win. Large multilingual systems benefit from shared knowledge across languages and from enormous training sets. But specialization gives a smaller team a plausible path to competitive accuracy without frontier-scale compute, and Haas-ja-v1.0 is a live test of that path. Its authors say it holds the top spot among the open models they compared, which, if it survives outside testing, supports the specialization argument. The practical angle for developers is simple. Transcribing a Japanese earnings call, a lecture, a customer support line, or a podcast all depend on the same core capability. When a model gets homophones and numerals right, downstream tasks such as summarization, search, and translation inherit cleaner input. Errors at the transcription stage propagate through every later step. ## Apache 2.0 and What an Open Release Means for Builders The Apache License 2.0 grants broad rights: commercial use, modification, distribution, and a patent grant from contributors. It asks for attribution and a copy of the license, and it does not require you to open-source your own product. For a startup or an enterprise legal team, that is close to the easiest license to approve. Open weights change the cost structure of transcription. With a hosted API, you pay per minute or per request and send audio to a third party. With a self-hosted model, you pay for hardware and operations, keep audio inside your own network, and control the upgrade schedule. Industries that handle sensitive recordings, such as healthcare, legal, and finance, often prefer the second arrangement. Japanese privacy and data residency expectations push in the same direction. A 615.8 million parameter model is also small enough to run on modest hardware compared with frontier language models. The sources do not publish hardware requirements, so this article makes no latency promises. Early community activity hints at fast adoption, though. A GitHub repository named [haas-mlx](https://github.com/matsuurajun/haas-mlx/blob/main/README.md) appeared in the sources, and its name suggests a port aimed at Apple’s MLX framework for Mac hardware. That is an inference from the repository name, not a confirmed feature list. If you want to try a local model workflow for the first time, our guide to [running an LLM locally in 12 steps](https://shattered.io/run-llm-locally-qwen3-8-27b-llama-cpp-12-steps-2026/) covers the general discipline of downloading weights, pinning versions, and measuring speed. The same habits apply to speech models. ## Open Japanese ASR vs Commercial APIs: Who Competes Haas-ja-v1.0 enters a market with three groups of competitors. Large multilingual open models handle Japanese as one language among many. Earlier Japanese-focused open efforts, such as Kotoba-Whisper, which was trained on a large subset of the ReazonSpeech dataset of Japanese television audio, proved that targeted training pays off. And commercial cloud transcription services sell convenience, scale, and support. | Approach | Strength | Trade-off | Where Haas-ja-v1.0 fits | |---|---|---|---| | Multilingual open model | One model for many languages | Japanese shares capacity with other languages | Targets Japanese only, per the authors | | Earlier Japanese open model | Proven fine-tuning on Japanese corpora | Older designs, authors’ comparison is the main evidence | Claims top accuracy among evaluated open models | | Commercial cloud API | No infrastructure, vendor support | Per-use fees, audio leaves your network | Model card calls its accuracy competitive at a small fraction of the cost | | Self-hosted Haas-ja-v1.0 | Apache 2.0, data stays local | You run and monitor it yourself | Open weights and inference code | AlphaSignal’s coverage leans hard into the cost story against commercial transcription APIs. Its headline carries a specific cost multiple. The primary announcement material available for this analysis establishes no price, so treat that multiple as the outlet’s framing and not as a verified figure. The model card’s own wording is more measured, and the next section quotes it. For perspective on how quickly open models are closing gaps with paid ones in other domains, see our comparison of [Muse Spark Open and Qwen3.8 on MCP-Atlas](https://shattered.io/muse-spark-open-vs-qwen3-8-mcp-atlas-89-2-vs-84-5-2026/). Speech is following the same arc that text models traveled over the past two years. ## What Matsushima and the Model Card Say Three statements from Matsushima appear in GIGAZINE’s coverage. Beyond the opening line quoted above, he described the model’s performance this way: “It achieves both high accuracy and inference efficiency, and is the most accurate open-weighted model we have evaluated.” The phrase “we have evaluated” is the key. It confirms the claim covers the authors’ comparison set, not every model in existence. He also invited readers to try it. In the same report he said, “Demos are also available, so please feel free to use it!” A Hugging Face Space for the model is listed among the sources, and its [README is public](https://huggingface.co/spaces/hugging-apps/third-intelligence-haas-ja-v1-0/raw/main/README.md). Testing on your own audio takes minutes and beats reading any benchmark table. The Hugging Face model card, written by Third Intelligence as the model authors, opens with this line: “Haas-ja-v1.0 is a Japanese speech recognition model built to be both fast and accurate.” It goes further on cost: “Its accuracy is competitive with major commercial speech-to-text APIs at a small fraction of their cost.” This is a first-party claim. The card does not name a dollar figure in the material available here, and no independent test in the sources confirms the comparison. Readers should hold it as a hypothesis worth checking. Taken together, the statements show a consistent and fairly modest positioning. The authors do not claim to have solved Japanese speech recognition. They claim a good point on the speed and accuracy curve, shipped openly, with enough documentation to reproduce the evaluation. That restraint makes the release easier to trust. ## Market Impact for Japanese Enterprises and Developers The immediate winners are teams that already need Japanese transcription at volume. Media companies with large archives, contact centers, education platforms, and meeting-notes products all pay for speech-to-text today. A credible open alternative gives them leverage in vendor negotiations even if they never switch. It also gives them a fallback if a provider changes pricing or terms. Cloud transcription vendors face a slower squeeze. Few customers will rip out a working integration over a benchmark. But new projects now have a free baseline to test against, and procurement teams will ask why the paid option costs what it does. Vendors that compete on workflow features, such as speaker labeling, redaction, and compliance tooling, will hold their position better than those selling raw transcription alone. For startups, the release lowers the barrier to voice products in Japan. A founder can prototype a transcription feature with open weights on a single machine before committing to usage-based bills. That matters most in the early stage, when volume is unpredictable and every fixed cost hurts. Our [AI and machine learning coverage](https://shattered.io/ai-machine-learning/) follows how these small, specialized releases reshape the economics of AI products. Hardware makers and cloud providers benefit indirectly. Efficient models that run on commodity GPUs, or on laptops, expand the places where voice AI can live. The sources confirm no hardware partnerships, so this remains a structural observation and not a news item. ## Historical Context: From Cloud APIs to Specialized Open Models For years, the best speech recognition lived behind commercial APIs. Open options existed, but they trailed on accuracy and demanded heavy engineering. The arrival of large open multilingual models shifted that balance, and community efforts then adapted those models to individual languages. In Japanese, projects built on the ReazonSpeech corpus showed that television-derived training data could lift accuracy substantially. The next step is what Haas-ja-v1.0 represents. Instead of [fine-tuning](https://shattered.io/category/ai-machine-learning/) a general model, a team builds a purpose-made architecture around efficiency and trains it for one language. NVIDIA’s Parakeet-TDT lineage supplied the template. Recent language model research supplied the attention variant for the prediction network. Third Intelligence assembled the pieces and published the result under a permissive license. The pattern echoes what happened in text. Giant general models drew the headlines, then smaller specialized and open models took large shares of real workloads because they were cheaper to run and easier to control. We saw a fresh example when [DeepSeek open-sourced its super-agent on October 31](https://shattered.io/deepseek-open-sources-super-agent-oct-31-2026/), one day after the Haas-ja-v1.0 release. Two open releases in two days, in different modalities, signals how routine open distribution has become. ## Risks and Open Questions Several questions remain open as of November 3, 2026. The first is independent replication. Nobody outside Third Intelligence has published a verified rerun of the benchmark in the sources reviewed. The second is how it handles messy audio. Real audio brings background noise, overlapping speakers, dialects, and telephone compression, and a pooled public benchmark may underrepresent them. The third is operational detail. The sources reviewed do not give hardware requirements, throughput numbers, or latency figures, so teams cannot yet plan capacity from documentation alone. The fourth is the training data. Licenses on speech corpora vary, and enterprises will want to understand what went into the model before they put it into a regulated workflow. The Apache 2.0 grant covers the weights and code, but data provenance is a separate question that the fact sheet does not answer. The fifth is long-term support. A v1.0 release from a research team does not promise maintenance. Enterprises weigh whether a model will receive fixes and successors. The version number implies later releases are planned, though no schedule is confirmed. Finally, the usual caveats about transcription apply. Even a low CER leaves errors, and the errors cluster on names, numbers, and technical terms, which are the words that matter most in contracts or medical notes. Human review stays necessary wherever a mistake carries real cost. ## Five Predictions for Japanese Speech AI These are this publication’s own forecasts, based on the release and past patterns. They are not statements from Third Intelligence or any named person. - **Community benchmarks will appear within weeks.** Open releases invite quick reruns, and Japanese developers have an active habit of publishing comparisons on their own audio. - **Ports and quantized builds will follow.** The haas-mlx repository is an early sign. Expect more runtimes and desktop tools to wrap the model. - **Rival Japanese labs will answer.** A public leaderboard position invites competitors to beat it, so expect new open Japanese ASR releases within months. - **Hybrid deployments will win in enterprises.** Many firms will run an open model for routine audio and keep a commercial API for hard cases or support guarantees. - **Specialization will spread to other languages.** If single-language models keep matching larger multilingual systems, teams will build similar models for Korean, Thai, and other languages with complex scripts. ## How to Evaluate Haas-ja-v1.0 on Your Own Audio Benchmarks tell you about someone else’s audio. Your own recordings tell you what matters. A fair test needs a reference transcript for a sample of your files, a consistent normalization step, and a scoring script. The open source jiwer library computes CER directly and works for Japanese because it operates on characters. ``` python import jiwer reference = "今日は会議の議事録を作成します" hypothesis = "今日は会議の議事録を作成します" # Apply the same normalization to both sides before scoring def normalize(text): return text.replace(" ", "").replace("　", "") score = jiwer.cer(normalize(reference), normalize(hypothesis)) print(f"CER: {score:.4f}") ``` Run the same script against each candidate system, including your current vendor, on the same files. Log the audio length and wall-clock time too, so you capture speed as well as accuracy. Split results by condition (clean speech, phone audio, noisy rooms) because an average hides the segments where a model struggles. Pay special attention to proper nouns, numerals, and domain jargon. If your business transcribes medical terms or product names, add a custom check for those words. And keep the normalization rules fixed across systems. Changing them between runs is the fastest way to produce a misleading comparison. ## The Takeaway Haas-ja-v1.0 gives Japanese-language builders a new open option with a clear architecture, a permissive license, and a first-party accuracy claim of 6.12% strict CER on a five-dataset benchmark. The confirmed record stops there. It does not establish a price, it does not prove a universal ranking, and it has not yet faced wide independent testing. That is still a meaningful release. A 615.8 million parameter model that the authors call the most accurate open model they evaluated, shipped with weights, code, and demos, lowers the cost of trying Japanese voice features. The sensible move for engineering teams is simple: run the demo, score it on your audio, and decide with your own numbers. ## FAQ **What is Haas-ja-v1.0?** It is a Japanese speech recognition model from Third Intelligence, designed for fast and accurate transcription. It was released on October 30, 2026, with open weights and inference code. **How accurate is it?** The authors report a 6.12% strict character error rate on a benchmark built from five public Japanese datasets. That is a first-party figure, and independent replication has not been reported in the sources reviewed. **Is it really the best open Japanese ASR model?** The model card says it was the most accurate open model the authors evaluated. That claim is limited to their comparison set and benchmark. No source supports a claim of being best on every test. **How big is the model and what is it built on?** It has 615.8 million parameters. The design follows NVIDIA’s Parakeet-TDT approach, with a FastConformer encoder and a Token-and-Duration Transducer decoder. The prediction network uses Kimi Delta Attention. **Can I use it in a commercial product?** The weights and inference code are released under the Apache License 2.0, which permits commercial use with attribution and license notices. Check the repository for the exact terms before you ship. **How much does it cost?** The available sources establish no price or subscription. The model is open, so your cost is the hardware and operations needed to run it. The model card describes its accuracy as competitive with commercial APIs at a small fraction of their cost, but gives no confirmed dollar figure in the material reviewed here. **Does it support languages besides Japanese?** The model is specialized for Japanese. The sources describe it as a Japanese speech recognition model and mention no other languages. **Where can I try it?** Matsushima said demos are available, and a Hugging Face Space for the model is listed among the public sources. The inference code is also released for local use.