YAML Metadata Warning:The pipeline tag "text2text-generation" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other

Charsiu G2P multilingual byT5-small 100 — TurboQuant TQ6 (native Metal)

A TurboQuant TQ6 bundle of the Charsiu multilingual byT5-small grapheme→IPA G2P model, built for native-Metal inference in the Bee streaming ASR app. It is the text→IPA phonemizer behind Bee's hotword correction (the sibling of the bearcove/zipa-large-crctc-ns-no-diacritics-780k-tq4 audio→IPA model).

What this is

Base model: charsiu/g2p_multilingual_byT5_small_100 (~300M params, byT5-small: d_model 1472, d_ff 3584, 12 enc / 4 dec, 6 heads, vocab 384, gated-gelu, byte-level tokenization).
Quantization: TurboQuant TQ6_1S — 6-bit, 32-value blocks with a Randomized Hadamard Transform front-end, dual half-block fp16 scales, and Lloyd-Max-optimal Gaussian centroids. The 129 Linear weights are quantized; the embedding tables (shared / *.embed_tokens), norms, and relative_attention_bias stay f32.
Runtime: consumed by helix-metal's g2p_metal::G2pModel — the quantized weights stay TQ6-resident in Metal (no dequant); split <name>.scales / <name>.qs tensors + an embedded turboquant.index.
Size: ~271 MB. Runs on-demand (phonemize when hotwords change, cache the IPA), so the footprint is transient.

Usage (input / output format)

byT5 byte-level tokenization, greedy decode, no beam search:

input  = byte+3 of  "<lang>: word"   (e.g. "<eng-us>: serde"; NO eos)
output = generated byte ids -> filter >= 3 -> (id - 3) as u8 -> UTF-8 IPA

Quality

Measured vs the fp32 model (pure quantization damage; eng-us, oracle = fp32): 3.7% WER / 0.61% CER — effectively transparent (== fp32 on the Charsiu dict), and byT5-small beats byT5-tiny by ~12 WER points. The TQ6-resident Metal path matches the dense forward (3.3% WER, 145/150 exact), confirming the kernel.

Provenance

Produced by hx g2p-pack --format tq6 from the upstream model. This is a Bee model bundle; the artifact layout is project-specific to Bee. Licensed MIT, matching the published sibling ZIPA bundle.

Downloads last month: 7

Safetensors

Model size

0.3B params

Tensor type

F32

Model tree for bearcove/g2p-multilingual-byT5-small-100-tq6

Base model

charsiu/g2p_multilingual_byT5_small_100

Finetuned

(2)

this model