Fine‑tuned Japanese HuBERT‑base for CTC phoneme recognition (v3), adding MeCab N‑best with pyopenjtalk‑plus post‑processing, a weighted loss combining CTC and MeCab costs, updated exclusion rules, and a new “ty” phoneme.
Description in other languages
Statistics
Statistics as of Nov 16, 2026
Hugging Face downloads (last 30 days)
2,000
Hugging Face likes
5
Listed
Sep 30, 2025
Downloads (30-day window) over time
Show as a table
| Month | DL (30d) |
|---|---|
| Dec 2025 | — |
| Jan 2026 | — |
| Feb 2026 | — |
| Mar 2026 | — |
| Apr 2026 | — |
| May 2026 | — |
| Jun 2026 | 2,000 |
| Jul 2026 | 1,000 |
| Aug 2026 | 1,000 |
| Sep 2026 | 5,000 |
| Oct 2026 | 3,000 |
| Nov 2026 | 2,000 |
Downloads over the 30 days before the last snapshot of each month.
Popular in the same category
- japanese-hubert-base-phoneme-ctc-v2prj-beatrice7kFine‑tuned Japanese phoneme CTC model based on rinna/japanese-hubert-base, released as a higher‑accuracy v2 that excludes sentences containing “日本/方/明日”, removes those with “十/0/0” whose pyopenjtalk-plus pronunciation includes “ジュッ”, and changes the “言う” pronunciation from “イウ” to “ユー”.
- japanese-hubert-base-phoneme-ctcprj-beatrice4kFine‑tuned Japanese phoneme CTC model based on rinna/japanese‑hubert‑base, trained with ReazonSpeech v2 data and pyopenjtalk‑plus labels, achieving improved accuracy in a new v2 release (prj-beatrice/japanese-hubert-base-phoneme-ctc-v2).
- wav2vec2-large-xlsr-53-japanesejonatasgrosman16MJapanese wav2vec‑2 XLSR‑53 fine‑tuned on Common Voice 6.1, CSS10, and JSUT, requiring 16 kHz audio and usable via HuggingSound or HuggingFace pipelines.
- wav2vec2-large-xlsr-japanese-hiraganavumichien153kA Japanese speech‑recognition model fine‑tuned from facebook/wav2vec2‑large‑xlsr‑53 on Common Voice and the JSUT corpus, optimized for 16 kHz audio input.
- anime-whisperlitagin33kAnime Whisper is a lightweight Japanese ASR model fine‑tuned on ~5,300 h of anime‑style dialogue that delivers low hallucination, rhythm‑aligned punctuation and accurate transcription of non‑verbal sounds and NSFW content, and must be run without an initial prompt.
- kotoba-whisper-v2.0kotoba-tech25kKotoba‑Whisper v2.0 is a Japanese ASR model distilled from OpenAI Whisper large‑v3, trained on 7.2 million ReazonSpeech clips, that runs 6.3× faster while matching the teacher’s CER/WER on in‑domain tests and includes stable‑ts/punctuation support and full training code on GitHub.
Resources in similar research fields
- Qwen3-ASR-1.7B-JAneosophie13kFine‑tuned Qwen3‑ASR‑1.7B for Japanese ASR, optimized to accurately transcribe proper nouns, organization and product names, kanji‑heavy and mixed Japanese/English technical terms.Shared: Speech Recognition・Speech & Audio in NLP
- kotoba-whisper-v2.0-fasterkotoba-tech13kKotoba Whisper v2.0 is converted to CTranslate2 format for use with CTranslate2 and faster‑whisper, providing installation, inference examples, Apple M2 benchmarking, and conversion instructions.Shared: Speech Recognition・Speech & Audio in NLP
- japanese-hubert-base-phoneme-ctc-v4prj-beatrice12kFine‑tuned Japanese Hubert‑Base for CTC phoneme recognition (v4) with updated sentence‑filtering rules, pronunciation adjustments, and a GPU switch to A6000, stopped training at 110k steps.Shared: Speech Recognition・Speech & Audio in NLP
- japanese-wav2vec2-base-rs35khreazon-research10kJapanese‑wav2vec2‑base‑rs35kh is a 96.7 M‑parameter wav2vec 2.0 Base model fine‑tuned on the ReazonSpeech v2.0 Japanese ASR corpus, reaching a 13.22 % CER, deployable with Hugging Face transformers, and released under an Apache 2.0 license.Shared: Speech Recognition・Speech & Audio in NLP
- reazonspeech-nemo-v2reazon-research4kreazonspeech-nemo-v2 is a 619‑M‑parameter Japanese long‑form ASR model built on an improved Fast‑Conformer with Linearly Scalable Attention, trained on the ReazonSpeech v2.0 corpus, offering multi‑hour inference via a subword RNN‑T decoder (3000‑token SentencePiece) and distributed under Apache 2.0.Shared: Speech Recognition・Speech & Audio in NLP
- parakeet-tdt-0.6b-ja-GGUFcstr3kA GGUF‑converted 0.6 B Japanese Parakeet TDT‑CTC model, usable with CrispASR’s CLI for TDT decoding (CTC fallback), delivers 6.4 % CER on JSUT with word‑level timestamps, available in a full 1.24 GB bit‑exact build and a ~470 MB Q4_K quantised variant (which degrades after ~8 tokens, so F16 is recommended).Shared: Speech Recognition・Speech & Audio in NLP