african-speech-id

CPU-friendly, fast language identification for 1,386 African languages.

The model uses a fast version of Omnilingual ASR to turn speech into text, and a classification head that reads that text and names the language. Both run on CPU, so a phone or a laptop is enough.

What is in this repository

path	what it is	size
300m/	the language-ID head	265 MB
asr/	the speech recogniser it reads from	365 MB

Both come from here, so there is one place to fetch from. AfricanSpeechId.load() pulls only the head; the recogniser is downloaded when you call AfricanSpeechId.download_recogniser() and not before.

The recogniser is Meta's Omnilingual ASR, Apache 2.0, mirrored unmodified; asr/LICENSE and asr/ATTRIBUTION.md travel with it.

Speed

The reason to use this rather than a large audio classifier. Measured on one Xeon Platinum 8558 core, int8, on held-out audio:

	throughput
classification head alone	0.28 ms per call, 3,600 per second
full pipeline, 1 thread	3.5x realtime
full pipeline, 4 threads	9.3x realtime
facebook/mms-lid-4017, same machine	1.4x realtime

A ten-second clip is identified in under three seconds on a single core. MMS-LID needs no recogniser, but it is 970 million parameters and 3.9 GB against this head's 265 MB, and it is slower end to end even counting the recogniser this depends on.

Results

Evaluated against facebook/mms-lid-4017 on identical clips, same 30-second cap, both systems scored in the label space each can produce:

test set	languages	clips	this model	MMS-LID-4017
Waxal	28	1,675	0.609	0.604
omniASR corpus (test split)	76	4,557	0.425	0.270
combined	104	6,232	0.474	0.360

Per language: ahead on 59, behind on 38, tied on 7.

Measured through the int8 recogniser in asr/, so these are the numbers you get rather than a best case. Transcribing with the fp32 build instead scores 0.478 -- the training transcripts were produced that way, and the mismatch costs a third of a point.

MMS-LID covers 4,017 languages and is strong on well-resourced ones. This model covers 1,386 and is built for the tail underneath them, which is what the omniASR test set is made of.

Read these numbers with two things in mind. They come from two evaluation sets chosen because they cover long-tail African languages; other sets exist and give different answers, and on FLEURS — 17 major languages — MMS-LID scores higher than this model does. And 104 of 1,386 languages are evaluated here at all: the remaining 1,282 are trained but unmeasured against any independent source.

Usage

import soundfile as sf
import sherpa_onnx
from african_speech_id import AfricanSpeechId

model, tokens = AfricanSpeechId.download_recogniser()
rec = sherpa_onnx.OfflineRecognizer.from_omnilingual_asr_ctc(model=model, tokens=tokens)
lid = AfricanSpeechId.load()

wav, sr = sf.read("clip.wav", dtype="float32")
s = rec.create_stream(); s.accept_waveform(sr, wav); rec.decode_stream(s)
print(lid.classify(s.result.text))

On device there is no Python: sherpa-onnx produces the transcript and the head runs in onnxruntime through a C API, with Kotlin and Swift bindings. See the repository.

How much audio to give it

Five seconds minimum, ten for the best result. Accuracy climbs steeply with the amount of speech and flattens after roughly ten seconds. Check that most of the audio is speech before transcribing — silence contributes nothing and inflates the apparent duration.

Training data

Three sources, budgeted per language rather than per source, at most three hours each:

source	languages	clips
GRN African Speech	1,019	235,212
YouVersion Bible audio	427	76,202
JW publication audio	218	32,054

343,468 clips in total. The budget is spread across sources on purpose: 360 languages have two or three different narrators rather than one. Single-narrator data is the central risk in this kind of corpus — an earlier experiment on acoustic embeddings reached 0.976 in domain and 0.108 out of it, having learned voices rather than languages.

All three sources are religious recordings, so the training domain is narrow. That is the main reason to expect worse performance on conversational speech than the figures above.

How it is built

A linear classifier over character n-grams of the transcript, with tf-idf folded into the ONNX graph. 50,000 features over 1,386 classes; a 300,000-feature version scored 0.4833 against 0.4779, which does not justify twelve times the size. Trained on 40-character windows rather than whole transcripts, so training matches how short utterances are served.

Limitations

Closed set. The head always names one of its 1,386 languages, including for speech in a language it has never seen. The top-1 minus top-2 margin gives a rejection signal.

Confidence values are small. A softmax over 1,386 classes spreads thin and a confident answer often reads about 0.03-0.05; use the margin, not the raw score. Labels are ISO 639-3 codes where the source corpus carried one and language names otherwise.

Dialects are not separable. Grebo-Chedepo against Grebo-Buah, Asante Twi against Twi: the recogniser normalises toward one orthography per language and the distinction does not survive. Dialect subsets are merged into their language.

Most languages are unmeasured. 104 of 1,386 have independent evaluation.

Narrow domain. All training audio is religious narration.

Transcribed with the November 2025 omniASR build. A February 2026 v2 exists and was not used; it changes orthographic conventions — restored diacritics, different word segmentation, Ethiopic script instead of Latin for some Ethiopian languages — so results here are a floor rather than a ceiling.

Licence

Apache-2.0.