aoiandroid/mms-lid-1024-coreml-8bit

MMS-LID 1024 (Core ML, 8-bit Palettized)

Core ML conversion of facebook/mms-lid-1024 for on-device speech language identification. This variant uses 8-bit palettization (k-means): smaller and more ANE-friendly than the float16 base, with minimal accuracy loss in practice.

• Source: facebook/mms-lid-1024
• Input: Raw 16 kHz mono waveform, fixed 10 seconds (160,000 samples), shape (1, 160000) float32
• Output: Logits shape (1, 1024); argmax → class index. Map to ISO 639-3 via labels.json or mms_lid_id2label.json

Contents

File	Description
mms_lid_8bit.mlpackage	Core ML model (8-bit palettized, ANE-friendly)
labels.json	Ordered list of 1024 ISO 639-3 language codes
mms_lid_id2label.json	Index → language code mapping

When to use this variant

• Prefer smaller size and faster / ANE-optimized inference while keeping accuracy close to the base model.
• On par with base in runtime tests for most languages; use base (float16) if you need maximum consistency with PyTorch reference.

Usage on iOS / macOS

Same as the base model: load the .mlpackage, feed 10 s of 16 kHz mono as input_values, take argmax of logits, and look up the language in labels.json. Pad/trim and chunking recommendations apply as in the base README.

Limitations

Same as base: fixed 10 s input, L2 accent misclassification, English ↔ Hawaiian/Maori confusion. Use chunking and confidence threshold where appropriate.

Mac smoke test (Core ML)

On-device smoke run: each file under INPUT/audio was resampled to 16 kHz mono float32, padded or trimmed to 160,000 samples (10 s), then passed to input_values; pred is ISO 639-3 from argmax(logits); conf is softmax mass on the predicted class (runner-side).

Note: Filenames are hints only (e.g. English.mp3 is not ground truth). Low conf or known MMS-LID confusions (e.g. English vs haw) may still appear.

Raw runner log

MMS-LID 1024 Core ML 8-bit — Mac smoke test
Model: https://huggingface.co/aoiandroid/mms-lid-1024-coreml-8bit
Model dir: $PROJECT_ROOT/Log/mms_lid_1024_8bit_mac_test/model_repo
Audio dir: $PROJECT_ROOT/INPUT/audio
Compiled temp: /var/folders/ky/nmbswxzs0s79wdxndfw1y6wh0000gn/T/model_repo.mlmodelc
Compute: MLComputeUnits(rawValue: 2)
Input: input_values  Output: logits
Labels: 1024
Host: ams-macbook-air.local  macOS: Version 26.3.1 (a) (Build 25D771280a)
English.mp3  pcm_samples=9054841  pred=haw  conf=0.2388  max_logit=7.5508  time_ms=1969.7
Euskara.mp3  pcm_samples=1865769  pred=hin  conf=0.3650  max_logit=7.7617  time_ms=427.3
Guaraní.mp3  pcm_samples=1682285  pred=grn  conf=0.9993  max_logit=14.7812  time_ms=419.7
Yorùbá.mp3  pcm_samples=1067049  pred=haw  conf=0.4466  max_logit=7.9453  time_ms=392.9
afrikaasns.mp3  pcm_samples=2387800  pred=nld  conf=0.9994  max_logit=15.0156  time_ms=453.2
arabic.mp3  pcm_samples=2060120  pred=ara  conf=0.9989  max_logit=14.3047  time_ms=436.0
bengali.m4a  pcm_samples=7836432  pred=ben  conf=0.9985  max_logit=14.5000  time_ms=629.3
chinese.mp3  pcm_samples=12904245  pred=cmn  conf=0.9993  max_logit=14.3438  time_ms=1184.4
isiZulu.mp3  pcm_samples=1396819  pred=heb  conf=0.5809  max_logit=8.2422  time_ms=403.6
kiswahili.mp3  pcm_samples=1888757  pred=swh  conf=0.9989  max_logit=14.2891  time_ms=442.2
korean.mp3  pcm_samples=2364395  pred=kor  conf=0.9995  max_logit=15.2031  time_ms=477.8
russinan.m4a  pcm_samples=15431029  pred=rus  conf=0.2780  max_logit=7.8711  time_ms=768.5
test.mp3  pcm_samples=274560  pred=jpn  conf=0.9984  max_logit=14.5391  time_ms=391.9
日本語.mp3  pcm_samples=1798234  pred=jpn  conf=0.9984  max_logit=14.5625  time_ms=471.3

License

CC-BY-NC-4.0 (inherited from facebook/mms-lid-1024).

Citation

@article{pratap2023mms,
  title={Scaling Speech Technology to 1,000+ Languages},
  author={Pratap, Vineel and others},
  journal={arXiv preprint arXiv:2305.13516},
  year={2023}
}