# ReazonSpeech: a package index, not the corpus it is named after > The repository holds the Python tooling for Japanese speech recognition, not the audio. Four ASR packages target four different frameworks with parameter counts from 120M to 619M, and the root pyproject.toml is a v1 compatibility shim marked for deletion. **reazon-research/ReazonSpeech** — Massive open Japanese speech corpus - Repository: https://github.com/reazon-research/ReazonSpeech - Website: https://research.reazon.jp/projects/ReazonSpeech/ - Stars: 404 · Forks: 36 - Language: Python - License: Apache-2.0 - Published: 2026-09-17 · Updated: 2026-09-17 · Language: en - Canonical page: https://hysenlabs.com/projects/reazon-research-reazonspeech ## What lives in the repository is tooling, and the corpus is elsewhere The first line of the documentation settles the most common misunderstanding. This repository provides access to the main user tooling of the ReazonSpeech and AVista projects. The corpus itself, described as the world's largest open Japanese speech corpus, is not in the tree, and neither is the audio data. What is in the tree is a small index. The top level holds a `.gitignore`, the licence, an `README.rst`, a `pyproject.toml`, a `pkg/` directory holding the individual packages, and a `colab/` directory. That last one is worth noting for anyone evaluating the project: there is notebook material in the repository, which is the fastest way to see whether a model runs before committing an environment to it. Two projects are covered. ReazonSpeech is the corpus and its recognition models. AVista is a separate project described as aiming at a new vista of human and robot interaction, focused on noise-robust multimodal speech recognition. The name is an acronym, and the documentation explains the derivation: Audio-Visual Speech Transcription and Alignment, chosen so the name contains the Latin words for bird and for view or sight. Installation is delegated. The instruction is that installation instructions live in each package's own readme, with the package list below as the index. There is no single install command at the root, which follows from the next thing about this repository. ## Four recognition packages, four frameworks, and the parameter counts do not track Three packages carry speech recognition models, and the reason there are three is that they wrap three separate toolchains. Each is named after the framework it needs, which is the clearest signal in the whole index. `reazonspeech.k2.asr` is the Kaldi line, built on sherpa-onnx, and is described as a next-generation model that is very fast and accurate. Its parameter count is 159M. `reazonspeech.nemo.asr` requires NVIDIA NeMo and implements FastConformer-RNNT, described as fast and accurate, at 619M parameters. `reazonspeech.espnet.asr` requires ESPnet and uses a Conformer-Transducer model at 120M parameters. Those three numbers are the useful part and they are not in an order anyone would guess. The NeMo model is roughly four times the size of the Kaldi one and five times the size of the ESPnet one, so the description that reads most favourably is attached to the largest model. The README gives no accuracy figures, so the size to quality relationship is not something the documentation resolves, and choosing between them means picking a framework first and reading the benchmark literature second. The remaining three packages are not competing recognisers. `reazonspeech.evaluation` is a set of tools for evaluating ReazonSpeech models and other speech recognition models, which is the piece that lets you make the comparison yourself. `reazonspeech.avsr` holds audio-visual speech models for AVista, including pretrained ones. `reazonspeech.espnet.oneseg` is a set of tools for analysing Japanese one-segment television streams, and it is described as the package to use to create a Japanese audio corpus. ## The bilingual model is framed for language detection, not for transcription The Kaldi package contains more than one model, and the second one has a different job from the first. Alongside the Japanese recogniser, there is a bilingual Japanese and English model, and the stated strength is that it is highly accurate at language detection in bilingual settings of the two languages. That is a narrower claim than bilingual transcription and worth keeping separate. A model tuned to decide which language a span of audio is in is solving a segmentation problem, and it is useful when your audio switches between Japanese and English mid-utterance, or when you are deciding which model to hand a segment to. The documentation does not claim it transcribes both languages well, and reading it that way would be an inference the text does not support. A development model is also named. For development purposes there is a model identified as ja-en-mls-5k, trained on 5k hours of ReazonSpeech and 5k hours of MLS English data each. Naming the training set sizes is more informative than it first appears, because it tells you the bilingual work is built on a balanced bilingual mixture rather than on a monolingual Japanese corpus with English appended. The wider context is that ReazonSpeech is the open Japanese corpus in this family, and other Japanese speech datasets that come up alongside it in the same research space, such as JTubeSpeech and LaboroTVSpeech, are separate projects rather than something this repository packages. ## The root pyproject.toml is a v1 shim that the repository wants to delete The file at the repository root is not the current package definition, and its own comments say so. It exists for compatibility with ReazonSpeech v1.x, with the stated purpose of ensuring that existing projects depending on the v1 API keep working without any mutual migration, and it carries a to-do line saying it should be deleted once people stop using v1.x. The version inside it is 1.1.1, which is a long way behind the v3.0.0 tag. So there are two version numbers in play in one repository, and which one you get depends on where you install from. That is the single most confusing thing about the layout, and it is worth resolving before you start, because a v1 install and a v3 install do not have the same package names. The mapping explains the rest. The setuptools configuration points the package directory at `pkg/_v1/src`, so installing the root project gives you the v1 source tree. The dependency list is the v1 dependency set: ginza and ja_ginza for Japanese text processing, soundfile and ffmpeg-python for audio, and what the comment calls the minimum set of modules required to use ESPnet, which is six, typeguard pinned to the 2 series, numpy held below 1.24, torch, torchaudio, espnet itself, and espnet_model_zoo installed straight from a git URL rather than from a package index. Those pins are the practical trap. numpy below 1.24 and typeguard on version 2 are both old enough to conflict with a current PyTorch environment, so installing the compatibility file into a modern project is likely to downgrade something you depend on. ## The audio-visual package follows the Hugging Face interface, which changes how you load it `reazonspeech.avsr` is the one package that does not require you to learn a project-specific loading convention. It follows the Hugging Face Transformers interface and supports the Auto classes for quick use, which means the familiar pattern of asking an Auto class for a pretrained model and getting a callable object back. That is a meaningful difference from the three ASR packages, which each require their own framework's loading path. In practice it means the audio-visual models can be tried in a notebook alongside other Transformers models without installing an extra inference stack, and the `colab/` directory at the repository root is consistent with that being the intended first contact. The models are pretrained ones for AVista, and AVista's own framing is noise-robust multimodal recognition for human and robot interaction. Noise robustness and multimodality are linked goals rather than separate features here: the point of adding a video channel is that a face and mouth movements carry information the audio channel loses when the room is loud, which is the situation a robot actually operates in. The documentation is not specific about the training data for the audio-visual models, the parameter counts, or the video requirements, so those are questions for the model cards rather than for the repository index. ## The version history crossed a major boundary in January 2026 Three tags are visible and they map onto a clean progression: v2.1.0 on 2024-08-01, v2.2.0 on 2025-08-05, and v3.0.0 on 2026-01-15. That is a little over eighteen months from the first of those to the last, with a thirteen month gap between the two v2 releases. The jump to 3.0.0 is recent enough that documentation written against v2 is worth checking, and the presence of the v1 compatibility file alongside a v3 tag means the project has been carrying migration for some time rather than treating it as a clean break. The default branch is master rather than main, which is worth knowing if you write automation against raw URLs. The last push to the repository is dated 2026-06-10, which is after the v3.0.0 tag. So the current state of the tree is ahead of the newest release, and if you are pinning to a tag you are pinning to January rather than to whatever is on the branch. Licensing is the simplest part. Apache License 2.0, copyright 2022 to 2025 Reazon Holdings. The organisation name in the copyright line is a detail that matters if you are considering the corpus for a commercial product, since it tells you who holds the rights on the tooling and, separately, who published the speech data. ## Conclusion This suits a speech researcher who wants a Japanese ASR baseline that is not Whisper, and who is willing to install a different deep learning framework per model, since the four ASR packages each require their own. It is a poor fit if you want the corpus itself, because the audio lives outside this repository, and a poor fit on Python 3.12 or newer given the version pins in the compatibility file. Before you install, check three things: which framework you already have in your environment, because sherpa-onnx, NVIDIA NeMo, and ESPnet are mutually awkward to keep side by side; whether you need the bilingual Japanese and English model, since that one is framed for language detection rather than transcription; and whether you are on a v1-era project, since the root install still points at the old package tree. The last push was 2026-06-10 and the newest tag is v3.0.0 from 2026-01-15. ## FAQ ### Does the ReazonSpeech repository contain the Japanese speech audio? No. The repository provides the user tooling for the ReazonSpeech and AVista projects, and the corpus itself is separate. What ships here is a set of Python packages under pkg/, including ASR models, audio-visual models, evaluation tools, and a tool for analysing Japanese television streams. ### Which speech recognition frameworks does ReazonSpeech support? Three, one per package. reazonspeech.k2.asr needs sherpa-onnx and uses a 159M parameter model, reazonspeech.nemo.asr needs NVIDIA NeMo with FastConformer-RNNT at 619M parameters, and reazonspeech.espnet.asr needs ESPnet with a Conformer-Transducer at 120M parameters. ### What is the Japanese and English model in ReazonSpeech for? The documentation describes it as highly accurate at language detection in bilingual settings of Japanese and English, rather than as a bilingual transcriber. A development model named ja-en-mls-5k is trained on 5k hours of ReazonSpeech and 5k hours of MLS English data each. ### Why does ReazonSpeech still ship a v1 compatibility package file? The root pyproject.toml exists so projects depending on the v1 API keep working without migration, and it points the package directory at pkg/_v1/src. It carries version 1.1.1 and a to-do line saying it should be deleted once v1.x usage stops, and it pins numpy below 1.24 and typeguard to version 2. ## Sources - [License: Apache-2.0](https://github.com/reazon-research/ReazonSpeech/blob/master/LICENSE) - [Project website](https://research.reazon.jp/projects/ReazonSpeech/) - [README](https://github.com/reazon-research/ReazonSpeech/blob/master/README.md) - [reazon-research/ReazonSpeech on GitHub](https://github.com/reazon-research/ReazonSpeech) - [Releases](https://github.com/reazon-research/ReazonSpeech/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/reazon-research-reazonspeech