codingSalafi/akoway-bark-3.0-pro-onnx

Underdog Bark 3.0 Pro — ONNX (for Akoway)

An ONNX conversion of ConwayResearch/Underdog-Bark-3.0-Pro by Conway Research, used by Akoway, a free offline voice keyboard for Windows. The app downloads these files on first run and runs them on the CPU with ONNX Runtime.

Files

file	size	what it does
preprocessor.onnx	1.2 MB	16 kHz waveform → normalized 128-bin log-mel features (same maths as ParakeetFeatureExtractor)
encoder.int8.onnx	839 MB	FastConformer encoder + encoder projector; MatMul weights int8, convolutions fp32
decoder.onnx	49 MB	LSTM prediction network, one token per call
joiner.onnx	21 MB	joint network → 8193 token logits + 5 duration logits (TDT)
tokens.txt	94 KB	SentencePiece vocabulary, blank last

Decoding is greedy TDT with 0.25 s of leading silence, as in Underdog's reference script. A reference implementation is in Akoway's tools/bark-onnx/compare_onnx.py.

Accuracy

74 LibriSpeech clean clips (73 with transcripts) plus one noisy dictation recording, CPU:

	WER
transformers reference (fp32)	4.00 %
these ONNX files	4.00 %

Quantizing every layer to int8 raised WER to 4.70 %, so only the MatMul weights are int8.

License and attribution

CC-BY-4.0, as the source. Underdog Bark 3.0 Pro is by Conway Research; it derives from nvidia/parakeet-tdt-0.6b-v3 (NVIDIA) and moondream/parakeet-ultra (Moondream). See LICENSE-NOTICE.txt. Changes in this repository: conversion to ONNX, audio preprocessing exported as a graph, and int8 quantization of the encoder's MatMul weights.