DengeVoice ASR Benchmark

Executive Summary

DengeVoice lets citizens report civic complaints by speaking naturally — mixing English with their local language, the way people actually talk. Choosing the right ASR engine for this is not cosmetic: it determines whether a citizen's complaint is captured correctly or silently mangled.

This report benchmarks three ASR engines — Sahara (Intron), Meta MMS (mms-1b-all), and NCAIR1's language-specific Whisper fine-tunes (referred to here as NATLAS) — against 30 code-switched audio clips per language across DengeVoice's four core languages: Igbo, Yoruba, Hausa, and Nigerian Pidgin, each mixed with English.

Headline result: Sahara achieved the lowest Word Error Rate (WER) in all four languages tested, confirming it as the right production ASR choice for DengeVoice.

---

Dataset & Code-Mixing Context

Testing audio came straight from intronhealth/AfriSwitch — a 54-hour public benchmark of real, conversational, code-switched speech across multiple African languages. From that pool, I pulled a sample of 30 clips per language to run the tests.

A quick look at the dataset stats highlights how different these languages behave in practice:

Language	Hours	Utterances	Avg. Switch Points	Total Switch Events	Code-Mixing Index (CMI)
Igbo	5.00	1,848	4.10	7,575	27.64
Yoruba	5.00	1,877	5.33	10,002	22.93
Hausa	5.00	1,515	4.00	6,053	13.09
Pidgin	4.56	1,801	4.29	7,719	30.15

What the numbers tell us: Pidgin has the highest Code-Mixing Index (30.15), meaning it involves the heaviest, most balanced back-and-forth mixing. Even with that complexity, Sahara hit its lowest error rate here (26.66% WER). On the flip side, Hausa has the lowest CMI (13.09) — sitting much closer to straightforward speech with just occasional English terms thrown in — where Sahara also performed very well (31.18% WER). Yoruba sits in the middle for mixing intensity, but it actually has the highest switch-point average (5.33), which lines up with how difficult it turned out to be for every model tested.

---

Per-Language Error Analysis

Igbo

Sahara’s mistakes were mostly substitutions (248) rather than insertions (18). That tells me it usually knew when someone was speaking and where the words went, even if it occasionally guessed the wrong vocabulary—which is way better than hallucinating text out of thin air. On the other hand, NCAIR1/Igbo-ASR racked up 298 insertions on the exact same clips, showing a tendency to over-generate words not actually found in the source audio when faced with messy conversation.

Yoruba

Every single model struggled the most here, pushing substitution counts way up (405 to 578). Yoruba's tonal diacritics and rapid switching clearly present a tough hurdle for current speech models across the board rather than exposing a flaw in just one system. It's a clear area where future fine-tuning is going to be necessary.

Hausa

Hausa gave Sahara one of its cleanest runs, keeping insertion errors down to just 53. NATLAS also bounced back significantly here compared to its Igbo and Yoruba scores, showing that open models perform much better when their training data aligns closer to the test distribution.

Pidgin

This yielded Sahara's strongest overall result at 26.66% WER. Because Nigerian Pidgin is English-lexified, other models had to rely on proxy adapters instead of purpose-built architectures. Sahara's success here comes down to having native support for the language pattern straight out of the box.

---

Reproducibility

If you want to run or check the pipeline, everything is sitting in scripts/:

Script	Purpose
prepare_data.py	Pulls test clips and ground truth for a language from AfriSwitch
transcribe_sahara.py	Hits the live Sahara API endpoint used by DengeVoice
transcribe_natlas.py	Runs NCAIR1's per-language Whisper fine-tunes locally
transcribe_mms.py	Runs Meta's mms-1b-all locally
score_wer.py	Computes WER, CER, and error breakdowns against ground truth

To run a test for any language:

python scripts/prepare_data.py --language <language> --n 30
python scripts/transcribe_sahara.py --language <language>
python scripts/transcribe_natlas.py --language <language>
python scripts/transcribe_mms.py --language <language>
python scripts/score_wer.py --language <language>