EdgeASR β€” Offline English Zipformer

This is the user-provided offline English recognizer from english_only_v9_6w_offline_no_delay_penalty/epoch-30-avg-9.

This is an offline, non-streaming model. Its non-causal encoder processes the complete recording. This package does not provide a streaming state or chunk-by-chunk inference API. The configuration sets causal=false, chunk_size=-1, and left_context_frames=-1.

The checkpoint, token table, and BPE model are unchanged. Its configuration is recorded in model_config.json. artifact_manifest.json records the exact checkpoint and inference-source hashes. The supplied icefall source retains its Apache 2.0 license in source/LICENSE. The weights use CC BY-NC 4.0, matching the author's published frankyoujian/EdgeASR model card.

FFASR inference

evaluator.py implements FFASR's evaluate(path: pathlib.Path) -> str interface. It loads the model once, uses the supplied icefall pretrained.py model and frontend, and performs greedy RNN-T decoding with at most one emitted symbol per encoder frame. Evaluation uses FP32 and batch size 1. It uses no external language model, prompts, reference transcripts, or dataset-specific decoding.

The frontend uses 16 kHz audio, 80-dimensional log-mel features, dither 0, snip_edges=False, and high_freq=-400, matching the supplied inference script. Audio with another sample rate is resampled to 16 kHz.

The pinned requirements target Linux x86-64, Python 3.12, and CUDA 12.4. For local inference from a downloaded copy of this repository:

python -m pip install -r requirements.txt
export FFASR_MODEL_DIR="$PWD"
python -c 'from pathlib import Path; from evaluator import evaluate; print(evaluate(Path("sample.wav")))'

For Hub-based loading, set FFASR_MODEL_ID=frankyoujian/EdgeASR and FFASR_MODEL_REVISION to the complete 40-character commit SHA. FFASR custom evaluation uses an immutable revision. The evaluator enforces offline configuration and returns the final formatted transcript.

Output formatting

<cap> capitalizes the preceding word; <upp> uppercases it. <.>, <,>, and <?> become ordinary punctuation. These formatting annotations are removed after their effects are applied. Formatting does not correct words or apply the benchmark's WER normalization.

Example:

we <cap> are from us <upp> <,> and they are from china <cap> <.> who <cap> are they from <?>
We are from US, and they are from China. Who are they from?

Results

No FFASR score is claimed in this model card. Any official score must come from the FFASR evaluation service after moderation and evaluation.

Training data

A detailed training-data disclosure has not been supplied with this checkpoint.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support