language:
- ja
pipeline_tag: automatic-speech-recognition
base_model:
- KeisukeMiyamoto/zipformer-cpt
- KeisukeMiyamoto/lfm2.5-230m-cpt
datasets:
- KeisukeMiyamoto/whisper-cluster-audio
- KeisukeMiyamoto/audio-transcript
- KeisukeMiyamoto/lambda-audio-courpus
metrics:
- cer
tags:
- japanese
- speech-to-text
- lambda-audio
- zipformer
- lfm2
- pytorch
model-index:
- name: LambdaAudio v6
results:
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
type: google/fleurs
name: FLEURS Japanese
config: ja_jp
split: test
metrics:
- type: cer
name: CER (%)
value: 6.05
LambdaAudio v6
LambdaAudio v6 is a Japanese automatic speech recognition (ASR) model that converts speech into Japanese text. It combines a Zipformer audio encoder with an LFM2.5 language decoder, both adapted to Japanese audio or transcripts before joint training.
On Japanese subset of Google FLEURS ASR benchmark, it achieves a character error rate (CER) close to Whisper large-v3-turbo (6.05% vs. 5.83%) while transcribing audio 2.3× faster.
Model details
| Field | Value |
|---|---|
| Developer | Keisuke Miyamoto |
| Task / language | Speech-to-text / Japanese (ja) |
| Model type | Encoder-decoder with an audio adapter and gated cross-attention |
| Model size | 350M parameters |
| Audio encoder | KeisukeMiyamoto/zipformer-cpt |
| Text decoder / tokenizer | KeisukeMiyamoto/lfm2.5-230m-cpt |
| Input | 16 kHz audio, up to 30 seconds per segment |
| Output | Japanese transcription text |
| Audio feature interval | 20 ms after the encoder; 40 ms after the adapter |
Architecture
The Zipformer encoder extracts features from 16 kHz audio. An audio adapter reduces the sequence length and projects these features into the decoder's hidden dimension. Gated cross-attention in the LFM2 full-attention blocks lets the decoder use audio features to predict the next text token.
Training data
| Stage | Dataset | Sources and labels |
|---|---|---|
| Encoder continued pretraining | Whisper Cluster Audio | Internet Archive, publicly available podcasts, and radio; 500 cluster labels derived from Whisper large-v3-turbo encoder features |
| Decoder continued pretraining | Japanese Audio Transcripts | Japanese transcripts from NHK archives and technology podcasts |
| Joint ASR training | Lambda Audio Corpus | NHK archives, technology podcasts, and ReazonSpeech |
Benchmarks
Each cell shows CER (%) / total transcription time (seconds). Lower values are better.
| Model | ReazonSpeech | Common Voice 25 Japanese | FLEURS Japanese |
|---|---|---|---|
| LambdaAudio v6 | 10.75 / 121.55 | 23.87 / 167.84 | 6.05 / 20.75 |
| Whisper large-v3-turbo | 11.42 / 354.04 | 26.40 / 592.25 | 5.83 / 47.82 |
| Kotoba-Whisper v2.2 | 11.70 / 322.53 | 22.89 / 541.39 | 7.35 / 46.15 |
| ReazonSpeech wav2vec2 | 14.01 / 49.35 | 31.43 / 68.22 | 15.04 / 9.22 |
| Parakeet tdt_ctc-0.6b-ja | 10.62 / 55.10 | 19.99 / 82.28 | 5.94 / 9.74 |
| Cohere Transcribe 03-2026 | 6.45 / 176.00 | 21.58 / 227.42 | 4.36 / 30.77 |
ReazonSpeech
Common Voice 25 Japanese
FLEURS Japanese
