mms-lid-1024-coreml-joint

MMS-LID 1024 (Core ML, 8-bit + INT8 LUT Joint)

Core ML conversion of facebook/mms-lid-1024 for on-device speech language identification. This variant uses 8-bit k-means palettization plus INT8 LUT joint optimization (iOS 18+): good balance of size, speed, and accuracy.

• 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

When to use this variant

• Target iOS 18+ and want a single quantized model with good accuracy/size trade-off.
• In runtime tests, 8bit-int8 may differ from base/8bit on a few languages (e.g. Euskara, Yorùbá, Russian); use base or 8bit if you need maximum agreement with PyTorch.

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. Requires iOS 18+ for full joint optimization support.

Limitations

Same as base: fixed 10 s input, L2 accent misclassification, English ↔ Hawaiian/Maori confusion. Slightly higher divergence from PyTorch than base/8bit on some files; use chunking and confidence threshold where appropriate.

<!-- BEGIN_MMS_LID_MAC_TEST -->

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.

<details> <summary>Raw runner log</summary>

MMS-LID 1024 Core ML — Mac smoke test
Model: https://huggingface.co/aoiandroid/mms-lid-1024-coreml-joint
Model dir: $PROJECT_ROOT/Log/mms_lid_1024_joint_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=0.2406  max_logit=7.3984  time_ms=1185.3
Euskara.mp3  pcm_samples=1865769  pred=hin  conf=0.3924  max_logit=8.8438  time_ms=412.6
Guaraní.mp3  pcm_samples=1682285  pred=grn  conf=0.9992  max_logit=14.5703  time_ms=417.0
Yorùbá.mp3  pcm_samples=1067049  pred=haw  conf=0.8309  max_logit=9.7266  time_ms=384.9
afrikaasns.mp3  pcm_samples=2387800  pred=nld  conf=0.9994  max_logit=14.9297  time_ms=445.0
arabic.mp3  pcm_samples=2060120  pred=ara  conf=0.9979  max_logit=13.6328  time_ms=431.3
bengali.m4a  pcm_samples=7836432  pred=ben  conf=0.9976  max_logit=14.0703  time_ms=589.6
chinese.mp3  pcm_samples=12904245  pred=cmn  conf=0.9993  max_logit=14.3359  time_ms=1314.8
isiZulu.mp3  pcm_samples=1396819  pred=heb  conf=0.3127  max_logit=7.0078  time_ms=400.7
kiswahili.mp3  pcm_samples=1888757  pred=swh  conf=0.9988  max_logit=14.1484  time_ms=416.3
korean.mp3  pcm_samples=2364395  pred=kor  conf=0.9994  max_logit=15.0938  time_ms=448.3
russinan.m4a  pcm_samples=15431029  pred=rus  conf=0.2187  max_logit=7.2305  time_ms=835.2
test.mp3  pcm_samples=274560  pred=jpn  conf=0.9987  max_logit=14.6172  time_ms=346.4
日本語.mp3  pcm_samples=1798234  pred=jpn  conf=0.9988  max_logit=14.6250  time_ms=484.4

</details> <!-- END_MMS_LID_MAC_TEST -->

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}
}