Atotti/Ja-FastConformer-CTC-25Hz-Streaming-100M

Japanese 25 Hz Causal FastConformer CTC

Japanese causal ASR model trained from scratch on ReazonSpeech using NVIDIA NeMo. The recommended artifact for inference is model.nemo. best.ckpt is also included for training resumption and inspection.

Model details

• Architecture: NeMo EncDecCTCModelBPE (FastConformer CTC)
• Parameters: 111,675,065
• Encoder: 17 layers, hidden size 512, 8 attention heads
• Audio: mono 16 kHz
• Encoder frame rate: 25 Hz (10 ms feature stride, 4x subsampling)
• Causality: causal downsampling and convolution, attention context [70, 0]
• Tokenizer: Japanese SentencePiece BPE, 3,000 tokens
• Training data: ReazonSpeech

Files

• model.nemo: self-contained inference artifact with weights, configuration, and tokenizer
• best.ckpt: Lightning/NeMo training checkpoint from epoch 12
• training/hparams.yaml: training-time model and optimizer configuration
• tokenizer/: unpacked SentencePiece tokenizer files
• eval/: evaluation metrics produced for the selected checkpoint

For portable inference, prefer model.nemo. The training checkpoint retains local training-data paths in its saved hyperparameters and may require path overrides when resuming on another machine.

Usage

from huggingface_hub import hf_hub_download
import nemo.collections.asr as nemo_asr

path = hf_hub_download(
    repo_id="Atotti/Ja-FastConformer-CTC-25Hz-Streaming-100M",
    filename="model.nemo",
)
model = nemo_asr.models.ASRModel.restore_from(path, map_location="cuda")
model.eval()

texts = model.transcribe(["audio.wav"])
print(texts[0])

The model itself is causal. NeMo's regular transcribe() interface processes a complete audio file; an application that emits partial results in real time must additionally use a chunked/cache-aware inference loop.

Evaluation

All normalized results use the project's ctc_ja_basic Japanese normalization.

Dataset	Samples	CER	Exact match
ReazonSpeech validation	10,000	0.1659	0.2967
JSUT basic5000	5,000	0.1905	0.1074
Common Voice 8.0 Japanese	4,483	0.2135	0.1383

Japanese transcripts are normalized without word-boundary spaces, so the recorded WER mostly behaves like an utterance-level error measure. CER is the more informative metric.

Integrity

• model.nemo SHA-256: 2ece5069990cf989338f386fe1cb4d4b0bab2f36c18257953184648812760b3e
• best.ckpt SHA-256: 3bda4b83250f7737eba29e1d828e8aa6060e6bfc58c06b41d9fc7476b5af1c13

Limitations

The model may produce omissions or substitutions, especially for short utterances, proper nouns, noisy speech, dialects, and domains unlike the training data. Do not use its output as the sole basis for safety-critical or high-impact decisions.