1 Commits
Author SHA1 Message Date
80e3019a74 feat(parakeet): add optional Underdog Bark 3.0 models, int8 and 4-bit (#1031)
* feat(parakeet): add optional Underdog Bark 3.0 int8 model

Adds bark-3.0 to the Parakeet model picker: Underdog Bark 3.0, an
English fine-tune of Parakeet TDT 0.6B v3, as the int8 sherpa-onnx
conversion published at velddev/Underdog-Bark-3.0-sherpa-onnx. It runs
on the existing sherpa-onnx recognizer, pinned to a revision with every
file checked against model_checksums.txt, like Orukeet (#840).

The conversion writes the unknown-word token where a quote mark was, so
the Parakeet transcript drops <unk>; the other models never emit it.

* docs(parakeet): say when to pick Bark 3.0; loosen its catalog test

The README now says Bark suits English dictation of conversational
speech and v3-european stays the choice for read or multilingual
speech. The test checks that bark-3.0 is offered rather than that it is
the last catalog entry, so a later model does not break it.

* test(parakeet): add docstrings to the Bark tests

* feat(parakeet): add optional 4-bit Underdog Bark 3.0 build

Adds bark-3.0-int4: the same Bark 3.0 weights quantized to 4-bit with
onnxruntime's HQQ quantizer, for machines short on memory. It needs
about half the RAM of the int8 build (0.64 GB against 1.28 GB peak) and
a 379 MiB download, at two to four times the CPU time (about 2x on
Intel, 3.7x on AMD by the publisher's measurement). The files are
published in velddev/Underdog-Bark-3.0-sherpa-onnx (Hugging Face PR #1),
and both Bark entries now pin its revision 3ae3af1, which holds both
builds; the int8 files are unchanged there.

Catalog entries gain an optional "precision" key, so the loader opens
encoder.int4.onnx and friends instead of assuming int8; every existing
model keeps the int8 default. The README section now covers both
builds in a few lines.

* test(parakeet): type the Bark test signatures

* docs(parakeet): name the Bark sherpa-onnx source and its pins

The README now says where both Bark builds download from (a pinned revision of
velddev/Underdog-Bark-3.0-sherpa-onnx), that it is an unofficial conversion, and
that each file is checked against its bundled SHA-256 pin, as the Orukeet
section does.

---------

Co-authored-by: Jatin <jatinkrmalik@gmail.com>
2026-11-02 12:31:08 -08:00