ai-ecoverse/Underdog-Bark-3.0-ONNX

Underdog Bark 3.0, ONNX

Underdog Bark 3.0 by Conway Research, an English speech recognition model fine-tuned from NVIDIA's parakeet-tdt-0.6b-v3, exported to ONNX for slicc-ortllama and plain onnxruntime. It was converted from model.bark at revision 85db2ad. That file is unpacked exactly with the source repo's own bark_loader.py, then exported with torch.onnx.export.

Files

Variant	Encoder	decoder_joint	Size
fp32	onnx/encoder_model.onnx	onnx/decoder_joint_model.onnx	2,509 MB
fp16	onnx/encoder_model_fp16.onnx	onnx/decoder_joint_model_fp16.onnx	1,291 MB
q8	onnx/encoder_model_quantized.onnx	onnx/decoder_joint_model_quantized.onnx	911 MB
q4f16	onnx/encoder_model_q4f16.onnx	onnx/decoder_joint_model_q4f16.onnx	524 MB

A variant is the encoder and decoder_joint graphs that share a suffix. The decoder_joint graph (prediction network and joint, ~70 MB) is fp32 in every variant.

Graphs

encoder_model*.onnx:

• Inputs: input_features f32 [B, T, 128] and attention_mask i64 [B, T].
• Outputs:
  • encoder_outputs f32 [B, T', 640], the FastConformer output after encoder_projector, 8x subsampled;
  • encoder_lengths i64 [B].

decoder_joint_model*.onnx computes one greedy TDT step:

• Inputs:
  • encoder_outputs f32 [B, 1, 640], one encoder frame;
  • input_ids i64 [B, 1], the last non-blank token (blank 8192 before the first);
  • state_h and state_c f32 [2, B, 640].
• Outputs:
  • token_logits f32 [B, 8193];
  • duration_logits f32 [B, 5], over durations [0, 1, 2, 3, 4];
  • state_h_out and state_c_out.

Greedy decoding, the way transformers' ParakeetForTDT.generate() does it:

1. Start at t = 0 with last = 8192 and zero states.
2. At frame t, take tok = argmax(token_logits) and dur = durations[argmax(duration_logits)].
3. If tok is blank, advance by max(dur, 1) and leave the state as it is.
4. Otherwise, emit tok, set last = tok, take state_*_out, and advance by dur (0 stays on the frame).
5. Stop when t >= encoder_lengths.

Text is the emitted SentencePiece pieces joined, with ▁ turned into a space. Consecutive repeats are not merged.

Features

The log-mel features are computed outside the graphs. preprocessor_config.json lists every parameter of transformers' ParakeetFeatureExtractor:

• 0.25 s of leading silence;
• pre-emphasis 0.97;
• centred STFT, n_fft 512, hop 160, symmetric Hann(400);
• power spectrum into 128 Slaney mel filters;
• log(x + 2^-24);
• per-feature mean and unbiased std over the valid frames, floor(samples / 160), with 1e-5 added to the std;
• padded frames set to 0.

export/tdt.py is a NumPy reference client. fixture/ holds one clip, its features from transformers, the mel filters, the encoder output, and the expected tokens, durations and text for every parity clip and variant.

Parity

The reference is transformers' feature extractor plus generate() on the unpacked fp32 weights. It was run on 20 English clips made with macOS say. The export was run on the same features, decoded greedily by tdt.py, with word error rate as slicc-ortllama computes it (case and punctuation folded). fp32 on tdt.py's own NumPy features also matches token for token. fp16 emits the same tokens as PyTorch on every clip, but on some x86 hosts a blank step or a duration can move by one frame.

	Exact tokens and durations	Same emitted tokens	WER vs PyTorch	WER vs script
PyTorch fp32	20/20	20/20	0	5.72%
fp32	20/20	20/20	0.00%	5.72%
fp16	19/20	20/20	0.00%	5.72%
q8	4/20	18/20	1.37%	5.05%
q4f16	0/20	13/20	2.05%	5.72%

License

CC-BY-4.0, the same as the source; see NOTICE. The model derives from:

• Underdog Bark 3.0 by Conway Research;
• nvidia/parakeet-tdt-0.6b-v3 by NVIDIA;
• moondream/parakeet-ultra by Moondream.

The conversion recipe is in export/, from ai-ecoverse/models.