CoolFace
Modelpublic

aufklarer/Omnilingual-ASR-CTC-1B-MLX-8bit

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face
0likes17downloads
Model Card

Omnilingual ASR — CTC 1B (MLX 8-bit)

MLX-compatible 8-bit quantization of Meta's Omnilingual ASR CTC-1B model for on-device inference on Apple Silicon (M1/M2/M3/M4). Prefer this variant when you need the smallest possible WER regression from fp32 and can afford an extra ~460 MB compared to the 4-bit build.

Omnilingual ASR is a wav2vec 2.0-style encoder-only model with a linear CTC head, trained by Meta for speech recognition across 1,600+ languages. The CTC variant is language-agnostic at inference time.

Model

Parameters1.01 B
FormatMLX safetensors (quantized linear layers + fp16 features)
Quantization8-bit per-group min-max, group size 64
Encoder layers48
Encoder dim1280
Attention heads20
FFN dim5120
Sample rate16 kHz (raw waveform input)
Frame rate50 fps
Max duration40 s
Languages1,600+
Vocabulary10,288 SentencePiece tokens

Files

FileSizeDescription
model.safetensors1006 MB8-bit quantized transformer weights + fp16 conv frontend
tokenizer.model1.2 MBSentencePiece tokenizer
config.json<1 KBArchitecture + quantization metadata

Architecture

Wav2Vec2FeatureExtractor (7-layer CNN, 320× downsample) → Linear 512→1280 → conv position encoder → 48× pre-norm Transformer encoder (dim 1280, 20 heads, ffn 5120) → LayerNorm → Linear CTC head (→ 10,288 tokens).

Performance

See the 4-bit variant for architecture notes and the 300M reference for FLEURS WER across en/fr/de/ar/hi. The 1B model is ~3× the encoder capacity and delivers correspondingly lower WER on low-resource languages.

Source

Links

License

Apache 2.0 (inherited from upstream).