mms-lid-1024-coreml

MMS-LID 1024 (Core ML, Float16)

Core ML conversion of facebook/mms-lid-1024 for on-device speech language identification on iOS 17+ and macOS. This is the base (float16) variant: best accuracy, no quantization.

• Source: facebook/mms-lid-1024 (Wav2Vec2ForSequenceClassification, ~1B params, 1024 languages)
• Input: Raw 16 kHz mono waveform, fixed 10 seconds (160,000 samples), shape (1, 160000) float32
• Output: Logits shape (1, 1024); argmax gives class index. Map index → ISO 639-3 using labels.json or mms_lid_id2label.json

Contents

Usage on iOS / macOS

1. Load mms_lid.mlpackage with Core ML (MLModel).
2. Ensure audio is 16 kHz mono float32. Pad shorter than 10 s with zeros, or trim longer to 10 s (first 160,000 samples).
3. Feed input: input_values = shape (1, 160000).
4. Get logits output, take argmax along the last dimension → predicted class index.
5. Look up the ISO 639-3 code in labels.json (array index) or mms_lid_id2label.json (key as string).

Recommendations: For long audio, split into ~6 s chunks, run LID per chunk, and use majority vote. Apply a confidence threshold (e.g. softmax max < 0.7 → treat as "unknown") to reduce false positives.

Limitations

• Fixed length: Model expects exactly 10 s of audio; pad or trim accordingly.
• L2 accent: Non-native-accented speech is often misclassified as the speaker's L1 (e.g. Japanese-accented English → Japanese).
• English ↔ Hawaiian/Maori: English is sometimes misclassified as Hawaiian (haw) or Maori (mri); use chunking and/or confidence threshold to mitigate.

<!-- 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
Model dir: $PROJECT_ROOT/Log/mms_lid_1024_coreml_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=0.2392  max_logit=7.5547  time_ms=1227.0
Euskara.mp3  pcm_samples=1865769  pred=hin  conf=0.4018  max_logit=7.9141  time_ms=451.6
Guaraní.mp3  pcm_samples=1682285  pred=grn  conf=0.9993  max_logit=14.8125  time_ms=427.8
Yorùbá.mp3  pcm_samples=1067049  pred=haw  conf=0.4383  max_logit=7.9219  time_ms=404.7
afrikaasns.mp3  pcm_samples=2387800  pred=nld  conf=0.9994  max_logit=15.0156  time_ms=483.1
arabic.mp3  pcm_samples=2060120  pred=ara  conf=0.9989  max_logit=14.3047  time_ms=470.2
bengali.m4a  pcm_samples=7836432  pred=ben  conf=0.9986  max_logit=14.5312  time_ms=616.0
chinese.mp3  pcm_samples=12904245  pred=cmn  conf=0.9993  max_logit=14.3516  time_ms=1400.2
isiZulu.mp3  pcm_samples=1396819  pred=heb  conf=0.5570  max_logit=8.1641  time_ms=460.3
kiswahili.mp3  pcm_samples=1888757  pred=swh  conf=0.9989  max_logit=14.2734  time_ms=503.7
korean.mp3  pcm_samples=2364395  pred=kor  conf=0.9995  max_logit=15.1953  time_ms=475.1
russinan.m4a  pcm_samples=15431029  pred=rus  conf=0.2830  max_logit=7.8672  time_ms=926.3
test.mp3  pcm_samples=274560  pred=jpn  conf=0.9984  max_logit=14.5234  time_ms=361.3
日本語.mp3  pcm_samples=1798234  pred=jpn  conf=0.9984  max_logit=14.5547  time_ms=527.0

</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}
}