Yoruba Transcription: Tones, Diacritics, and the Path Ahead

Yoruba is one of the most-spoken languages in Africa, with roughly 50 million speakers across southwestern Nigeria, Benin, Togo, and a substantial diaspora in the UK, US, and Brazil. The demand for transcription is real: Yoruba-language Nollywood channels now generate millions of views per day on YouTube, and podcasts in Yoruba have multiplied across every major platform. Yet as of mid-2026, Yoruba transcription sits in an uncomfortable position: technically in scope for major ASR engines, but not reliably accurate enough for production use.

This is an honest account of why, where research has arrived, and what you can actually do today.

The Three-Tone System Is Not a Footnote

Yoruba is tonal in a way that directly determines meaning, not just pronunciation. The language uses three register tones:

• High tone, marked with an acute accent: á
• Low tone, marked with a grave accent: à
• Mid tone, left unmarked

A classic illustration: ọ́kọ̀ (high-low) means "husband," ọ̀kọ̀ (low-low) means "hoe," and ọkọ̀ (mid-low) means "vehicle." The spoken distinction is pitch. The written distinction is diacritics. Remove the marks and you collapse meaning.

Beyond tone marks, standard Yoruba uses subdot diacritics for open vowels and a palato-alveolar consonant: ẹ (open-e, as in "pet"), ọ (open-o, as in "off"), and ṣ (similar to English "sh"). Tone marks layer on top of these: ẹ́, ẹ̀, ọ́, ọ̀. The orthography was standardized at a 1875 Lagos conference organized by missionaries and local scholars, and the standard has held.

A production-grade Yoruba transcript must preserve every diacritic. A transcript without them is technically readable but semantically lossy, the equivalent of stripping punctuation from English legal text.

Why ASR Struggles With Tone Diacritics Specifically

The problem is not just "low-resource." It is architectural.

A 2026 study published at ACL found that current transformer-based ASR models actually produce lower error rates on Yoruba text that has tone marks removed than on fully diacritized text. The authors attribute this to three factors: tokenization behavior (most tokenizers were designed for European scripts and split diacritized characters unpredictably), insufficient representation of tonal acoustic cues in the pre-training data, and the absence of tone modeling in the underlying model objectives.

This is a counterintuitive finding. It means the models are not simply failing to mark tones, they are being actively confused by the presence of marks in the reference text during training. Solving Yoruba transcription properly requires rethinking how tonal diacritics are tokenized, not just adding more Yoruba audio.

For an explanation of why this pattern appears across multiple low-resource tonal languages, see why AI struggles with low-resource languages.

Where the Numbers Stand

The NaijaVoices dataset paper (Interspeech 2025) provides the clearest public benchmark for Yoruba ASR on unfine-tuned models:

WER above 100% means the model produces more word errors than there are reference words, usually from hallucinated text or repeated phrases. Whisper Large at 106% baseline is essentially producing noise on clean Yoruba audio.

The fine-tuning results are where things get genuinely interesting. Fine-tuning Whisper Small on NaijaVoices drops the error rate by roughly 260 percentage points. That is dramatic improvement, though 58% WER is still far from the sub-15% range typical of production-ready English. The gap between "dramatically better" and "usable without human review" remains large.

For context on how these error rates compare to other languages and services, see transcription accuracy explained.

Dialect Fragmentation Makes It Worse

Standard Yoruba ASR, bad as it is, still outperforms Yoruba dialects by a measurable margin.

Research testing MMS and SeamlessM4T in zero-shot settings found an average 12-point WER gap between standard Yoruba and regional dialects. Specific numbers: standard Yoruba at 72.50 WER, Ifẹ̀ dialect at 85.38, Ìlàjẹ dialect at 83.79.

Yoruba has at least five regional groupings, Northwest (Oyo, Ogun), Northeast (Kwara), Central, Southwest, and Southeast, each with phonological, lexical, and consonant differences. The Ijebu dialect, spoken around Ogun State, has documented consonant alternations and vowel distinctions that do not appear in the standard written form. A model trained on news-reading speech in Lagos will not handle an interview recorded in Ekiti or Ondo without substantial accuracy loss.

When dialect-adaptive fine-tuning was applied with as few as 800 training instances per dialect, the researchers achieved roughly a 20-point WER reduction. Small data quantities can move the needle, but only if someone first collects and labels dialect-specific audio.

Code-Switching in Lagos Speech

Urban Lagos Yoruba does not stay in one language per sentence. Speakers move mid-phrase between Yoruba, English, and Nigerian Pidgin. A sentence might open in Yoruba, insert an English technical term, and close with a Pidgin expression, all grammatically coherent to the speaker.

A 2025 ACL paper on Yoruba-English code-switching found that fine-tuning monolingual ASR models on the YECS Corpus (120 hours of naturally produced code-switched speech from 140 bilingual speakers) achieved over 20% WER reduction on code-switched audio while maintaining competitive performance on standard Yoruba. Monolingual fine-tuned models were computationally cheaper than multilingual alternatives and still closed most of the gap.

The practical takeaway: a generic multilingual model will partially handle the English segments of code-switched speech but fumble the Yoruba portions and the transitions. Content recorded in formal, single-speaker, diacritized Yoruba will transcribe better than conversational Lagos speech, not because Lagos Yoruba is "incorrect," but because ASR models are trained on the former and not the latter.

For more on how code-switching affects accuracy across languages, see transcription for non-native English speakers.

Community Dataset Efforts: What Exists

The resource situation for Yoruba is better than it was five years ago. Several concrete datasets are now publicly available:

ÌròyìnSpeech (LREC 2024): 42 hours of fully diacritized Yoruba speech from 80 volunteer speakers, drawn from news and creative writing. Released under CC-BY-4.0, meaning it can be used for commercial model training. This is notable because the diacritized reference text makes it useful for studying tone marking specifically.

NaijaVoices (Interspeech 2025): 606.94 hours of Yoruba speech from 1,768 speakers, part of a 1,800-hour dataset covering multiple Nigerian languages. This is currently the largest open Yoruba speech corpus. The diversity of speakers is a meaningful step toward dialect coverage.

YECS Corpus: 120 hours of high-quality Yoruba-English code-switched speech, 99,930 validated audio-text pairs, from 140 demographically diverse bilingual speakers. Specifically designed for the code-switching problem.

Masakhane (the grassroots African NLP community, 1,000-plus participants across 30 African countries) has a stated 2025 goal of collecting 500 hours of speech for five underrepresented languages. Whether Yoruba is in that cohort is not publicly confirmed as of this writing, but the organisation has previously supported Yoruba NLP work through its benchmark and dataset initiatives.

The cumulative picture: over 700 hours of open Yoruba speech now exists across multiple collections. That is a workable foundation for fine-tuning. The constraint is not absence of data any more, it is the alignment between audio and fully diacritized reference transcripts, and the absence of large-scale dialect-balanced coverage.

What Production-Grade Yoruba Will Actually Require

Getting Yoruba to the 85-90% accuracy range needed for practical, lightly-edited transcription requires all of the following happening together:

1. Tone-aware tokenization. Current models must be retrained or fine-tuned with tokenizers that treat diacritized Yoruba characters as single units, not as base character plus combining mark. This is a pre-training problem, not just a fine-tuning patch.

2. Diacritized reference data at scale. News-domain data (what ÌròyìnSpeech provides) is a start. Conversational and dialect-diverse data at scale is what closes the gap between lab performance and real-world audio.

3. Code-switching handling baked in. For Lagos content especially, a model that treats Yoruba-English transitions as an error to minimize will always underperform on the actual material researchers and journalists need transcribed.

4. Dialect coverage. A 12-point WER penalty for dialects means a model that is mediocre on standard Yoruba is actually bad on regional speech. Broader speaker diversity in training data is necessary, not optional.

The 2026 research landscape suggests all four areas are being worked on simultaneously, which is a better situation than three years ago when the dataset foundation itself was the bottleneck.

Comparing What Exists Today

For related comparison context, see best speech-to-text APIs in 2026.

Practical Recommendations for 2026

If you need to transcribe Yoruba audio now:

For research and drafting: Submit with language code "yo" via an API that routes to Whisper or a fine-tuned checkpoint. The output will be partially correct Yoruba text, often with missing tone marks and some hallucinated segments. Budget a review pass with a fluent Yoruba speaker, particularly for any proper noun, place name, or term where tone changes meaning.

For formal, legal, or broadcast use: AI-only Yoruba transcription is not there yet. Human transcription with a certified Yoruba transcriber remains the accurate path. Use AI output as a rough first draft to reduce the transcriber's workload, not as a finished product.

For West African content where Yoruba is not strictly required: Hausa transcription operates at production accuracy today. Many Nigerian content creators work across both languages, and for content where the language can flex, Hausa is the cleaner current option.

My read on this: Yoruba has crossed the threshold from "the model does not know this language exists" to "the model produces something recognizable." The gap between that and production-ready is still substantial. The research momentum and dataset volume in 2025-2026 is the best signal yet that this gap will close, but the honest answer today is that any Yoruba transcript produced by an AI engine needs a knowledgeable human in the loop.

If you just need a clean transcript without a meeting bot or complex integrations, ConvertAudioToText supports Yoruba via API and gives you the raw output to edit. It is a starting point, not a finished solution for this language.

FAQ

Does Whisper support Yoruba?

Whisper includes Yoruba as language code "yo" and has done so since large-v3. But support in name is not support in practice. Baseline WER on Whisper Large for Yoruba is above 100%, meaning more words are wrong than right before any fine-tuning. After fine-tuning on the NaijaVoices dataset, Whisper Small drops to about 58% WER, which is better but still far from broadcast or legal quality.

Why does Yoruba break speech recognition more than other African languages?

Three overlapping reasons. First, Yoruba is a three-tone language where pitch changes word meaning, and current ASR tokenizers were not designed around diacritics. A 2026 ACL study found models actually produce lower error rates on tone-stripped text than on fully diacritized text, because the diacritics themselves break tokenization. Second, Yoruba has at least five regional dialect groupings with measurable phonological differences; a model trained on standard Yoruba shows roughly a 12-point WER gap on dialects like Ifẹ̀ and Ìlàjẹ. Third, urban Lagos speech code-switches heavily with English mid-sentence, and most models handle either language or the other, not the blend.

What community datasets exist for Yoruba speech?

Several verified ones exist as of mid-2026. ÌròyìnSpeech (published at LREC 2024) contains 42 hours of fully diacritized Yoruba from 80 volunteer speakers, released under CC-BY-4.0. NaijaVoices (Interspeech 2025) adds 606 hours of Yoruba speech with 1,768 speakers, the largest open set to date. The YECS Corpus contributes 120 hours of naturally produced Yoruba-English code-switched speech from 140 bilingual speakers. Masakhane, the community NLP organisation, has a 2025 goal to collect 500 hours for five underrepresented African languages, with Yoruba as a target.

Should I use Yoruba transcription tools now or wait?

For casual research or content where a rough draft plus manual editing is acceptable, using Whisper with language code "yo" is a viable starting point today. Plan a full editing pass by a fluent speaker, especially for tone marks and proper nouns. For broadcast, legal, or formal publishing workflows, the accuracy is not there yet. If your content allows it, Hausa transcription is currently production-ready at higher accuracy and covers significant overlap in West African content workflows.

Sources

• Tone in Yoruba ASR: Evaluating the Impact of Tone Recognition on Transformer-Based ASR Models (ACL 2026)
• ÒWE-Voice: Evaluation of Monolingual and Multilingual ASR Using Yoruba Proverb Speech (AfricaNLP 2026)
• Voices Unheard: NLP Resources and Models for Yorùbá Regional Dialects (EMNLP 2024)
• The NaijaVoices Dataset (Interspeech 2025)
• ÌròyìnSpeech: A Multi-purpose Yorùbá Speech Corpus (LREC 2024)
• Fine-Tuning Monolingual ASR for Yoruba-English Code-Switching (CALCS 2025)
• Automatic Speech Recognition for African Low-Resource Languages: A Systematic Review (2025)
• Masakhane community
• Yoruba language, Wikipedia
• OpenAI Whisper tokenizer, GitHub