Modelpublic

TechHara/timestamps-asr-canary-ctc-mlx-q8

sourceHugging Facecc-by-4.0updated 4mo agoView on Hugging Face

0likes34downloads

Model Card

canary-1b-v2 — timestampsasrmodel — MLX q8

Quantized MLX port of the timestamps_asr_model bundled inside nvidia/canary-1b-v2. It is the dedicated CTC forced-alignment head paired with canary; vocabulary is byte-identical to canary's sentencepiece (16 384 pieces).


Architecture	24-layer Conformer encoder + 1×1 CTC head
Params	626 M
Mel features	128
d_model	1024
CTC vocab	16 384 + blank
Subsampling	8× (dw_striding)
Self-attention	full rel-pos
Quantization	bits=8, group_size=64 (MLX default)
Source	extracted from `canary-1b-v2.nemo`

Files

File	Size
`config.json`	preprocessor + encoder + CTC config (with inline vocab)
`model.safetensors`	q8 weights + bf16 scales/biases
`tokenizer.model`	canary's sentencepiece (1:1 copy)

License

CC-BY-4.0 — inherited from nvidia/canary-1b-v2.