Modelpublic

bearcove/zipa-large-crctc-ns-no-diacritics-780k-tq4

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes7downloads

Model Card

ZIPA Large CR-CTC NS No-Diacritics 780k — TurboQuant TQ4 (native Metal)

A TurboQuant TQ4 bundle of the ZIPA-large (Zipformer, audio→IPA CTC) model, built for native-Metal inference in the Bee streaming ASR app. It is the audio→IPA phonemizer behind Bee's live hotword correction.

What this is

—Base model: `anyspeech/zipa-large-crctc-ns-no-diacritics-780k` (300M params, Zipformer CR-CTC non-streaming, no diacritics).
—Quantization: TurboQuant TQ4_1S — 4-bit, 32-value blocks with a Randomized Hadamard Transform front-end, dual half-block fp16 scales, and Lloyd-Max-optimal Gaussian centroids. The 410 Linear weights are quantized; convs / norms / biases / bypass scales / linear_pos stay f32.
—Runtime: consumed by helix-metal's zipa_metal::ZipaModel — the quantized weights stay TQ4-resident in Metal (no dequant), and the split <name>.scales / <name>.qs tensors are 16 KiB page-aligned so they're aliased zero-copy from the mmap via newBufferWithBytesNoCopy.
—Size: ~204 MB (≈180 MB data + page-alignment padding). On-disk == resident.

Layout

model.safetensors   bundle tensors + embedded `turboquant.index` (header metadata):
                      <name>.scales / <name>.qs   (u8, split TQ4 blocks) per Linear
                      conv2d (OHWI) / norms / biases / bypass / downsample (f32)
tokens.txt          CTC phone vocabulary (id → IPA unit)
config.json         architecture + quantization metadata

Quality

Validated against the upstream fp32 ONNX oracle (feeding identical fbank into both), greedy phone-sequence metric (what the hotword matcher consumes):

	frame-argmax	greedy phone-seq
upstream fp32 (Metal vs ONNX)	97.2%	99.2%
this TQ4 bundle	97.6%	100.0%

TQ4 reproduces the oracle's greedy phone sequence exactly on the validation utterance — quantization is effectively transparent for this model.

Provenance

Produced by hx zipa-pack --format tq4 from the upstream averaged checkpoint. This is a Bee model bundle; the artifact layout is project-specific to Bee. Licensed MIT, matching the sibling `bearcove/zipa-small-crctc-ns-no-diacritics-700k-mlx-q8` (same upstream ZIPA repository license).