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

VisualEars FastConformer Persian ASR LiteRT FP

Reference LiteRT/TFLite fixed-length acoustic CTC-core export of Reza2kn/visualears-fastconformer-fa-full-ab.

Artifact

• Format: LiteRT/TFLite FP fixed-length acoustic CTC-core export
• Quantization/conversion: FP LiteRT/TFLite conversion reference; no 4-bit quantization
• Runtime validation: LiteRT/TFLite XNNPACK CPU
• Size: 437 MB .tflite

Validation

Two different parity checks are available:

Check	Result
Frame-level CTC argmax parity vs PyTorch	100.00%
Greedy CTC transcript parity vs PyTorch on the same 16 calibration items	100.00% / 16 of 16

Transcript parity is computed after greedy CTC collapse with the model tokenizer. This is still a fixed-core feature-to-logits check, not a full raw-audio pipeline evaluation.

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

The LiteRT export required folding fixed positional slice constants and removing all-valid pre-encoder mask multiplies.