ReazonSpeech
Japan's largest open speech corpus and the 619M-parameter FastConformer model trained on 35,000 hours of Japanese TV audio.
About
ReazonSpeech is Japan's open speech answer to the multilingual giants: a continuously growing corpus of Japanese TV audio, now around 35,000 hours in v2.0, released freely for commercial use. Reazon Holdings trains its own models on it, with packages for sherpa-onnx (a 159M-parameter Zipformer with a bilingual Japanese-English variant), NVIDIA NeMo (a 619M FastConformer-RNNT), and ESPnet.
The original paper reported 8.23% character error rate on the JSUT basic5000 set and 9.93% on Common Voice v8.0, on par with Whisper large-v2 at the time. The v2 NeMo model handles long-form Japanese audio up to several hours, and everything is Apache-2.0. For Japanese-language products that want open weights and no per-minute fees, this is the strongest domestic option.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- 8.23% CER on JSUT basic5000, 9.93% on Common Voice v8.0 (paper-reported, ReazonSpeech v1)
- v2 corpus grew to ~35,000 hours of Japanese audio
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Apache-2.0 license and the 8.23% / 9.93% CER figures from the ReazonSpeech paper; 619M-parameter FastConformer v2 model and 159M sherpa-onnx variant confirmed on Hugging Face and the project repo. Strengths: the largest open Japanese speech corpus, commercially usable, multiple maintained model flavors. Docked for: (1) benchmark figures are from the v1 paper and are provider-run, with no independent eval of the v2 models found; (2) Japanese-focused, so multilingual needs go elsewhere. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · v2.0 corpus ~35k hours Japanese TV audio; 619M FastConformer-RNNT + 159M sherpa-onnx Zipformer (bilingual ja-en variant)
- · Paper CER: 8.23% JSUT basic5000, 9.93% Common Voice v8.0 (v1, provider-reported)