Modelpublic

Reza2kn/visualears-fastconformer-fa-full-ab-litert-w4

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes6downloads

Model Card

VisualEars FastConformer Persian ASR LiteRT W4

LiteRT/TFLite weight-only 4-bit export of `Reza2kn/visualears-fastconformer-fa-full-ab`.

Artifact

—Format: LiteRT/TFLite weight-only 4-bit fixed acoustic CTC-core export
—Quantization/conversion: AI Edge Quantizer weight-only 4-bit, selected FULLY_CONNECTED ops, regex Wrapper;[2-9], channelwise
—Runtime validation: LiteRT/TFLite XNNPACK CPU
—Size: 357 MB, 81.7% of LiteRT FP source

Validation

This artifact passed the earlier frame-level argmax gate, but it does not pass transcript parity.

Check	Result
Frame-level CTC argmax parity vs PyTorch	98.23%
Greedy CTC transcript parity vs PyTorch on 16 calibration items	37.5% / 6 of 16
Greedy CTC transcript parity vs LiteRT FP on 16 calibration items	37.5% / 6 of 16

The earlier 98.23% number is frame-level CTC argmax token agreement, not "98% same final transcripts." For transcript-parity-sensitive use, this W4 artifact should be treated as failed.

Usage Boundary

This export takes precomputed log-mel features as processed_signal; it is not a full raw-audio-to-text pipeline by itself.

Notes

Best compressed LiteRT 4-bit attempt so far, but not a transcript-parity artifact. All-FC W4 compressed to 28.0% but failed at 96.51% frame argmax parity; safe FC+conv failed at 97.51% frame argmax parity.