Modelpublic

Alkd/TEA-ASR-1.1-MLX-4bit

sourceHugging Facemitupdated 2mo agoView on Hugging Face

2likes78downloads

Model Card

TEA-ASR-1.1-MLX-4bit

MLX conversion of JacobLinCool/TEA-ASR-1.1 (MIT), a Taiwan-Mandarin ASR model adapted from Qwen/Qwen3-ASR-1.7B that emits Traditional Chinese with Taiwan vocabulary natively and handles Mandarin-English code-switching. All credit for the model goes to the TEA-ASR authors.

Converted with mlx-audio v0.4.5 using mixed affine quantization (the MLX equivalent of the GGUF Q4KM + mmproj-Q8_0 recipe).

Quantization recipe

Component	Quantization
Text decoder (+ embeddings)	4-bit, group size 64, affine
Audio tower (AuT encoder + projector) linear layers	8-bit, group size 64, affine
Conv / norm / positional layers	float (MLX has no quantized conv)

Per-layer bit widths are recorded in config.json under quantization[<layer path>] (standard MLX per-layer overrides): the audio tower linear layers carry {"bits": 8, "group_size": 64} entries while the global default is 4-bit / group size 64.

Usage (Python, mlx-audio)

Note: mlx-audio's stock qwen3_asr loader hardcodes "never quantize the audio tower" and checks that before the per-layer config overrides, so loading this model needs a one-line predicate override:

python

from mlx_audio.stt.models.qwen3_asr import qwen3_asr
qwen3_asr.Qwen3ASRModel.model_quant_predicate = lambda self, p, m: True

from mlx_audio.stt.utils import load_model
model = load_model("Alkd/TEA-ASR-1.1-MLX-4bit")
result = model.generate("audio.wav", language="Chinese")
print(result.text)

Loaders that follow the per-layer quantization dict in config.json (mlx-lm-style loading, custom MLX Swift loaders) need no override.

Intended use

Built for the Sumi macOS app (on-device Mandarin voice-to-text) via mlx-audio / mlx-audio-swift.

License

MIT (same as the base checkpoint). The underlying Qwen3-ASR weights remain subject to Apache-2.0 attribution terms.