AfriSpeech-200: Pan-African Accented Speech Dataset for Clinical and General Domain ASR

Abstract Africa has a very poor doctor-to-patient ratio. At very busy clinics, doctors could see 30+ patients per day—a heavy patient burden compared with developed countries—but productivity tools such as clinical automatic speech recognition (ASR) are lacking for these overworked clinicians. However, clinical ASR is mature, even ubiquitous, in developed nations, and clinician-reported performance of commercial clinical ASR systems is generally satisfactory. Furthermore, the recent performance of general domain ASR is approaching human accuracy. However, several gaps exist. Several publications have highlighted racial bias with speech-to-text algorithms and performance on minority accents lags significantly. To our knowledge, there is no publicly available research or benchmark on accented African clinical ASR, and speech data is non-existent for the majority of African accents. We release AfriSpeech, 200hrs of Pan-African English speech, 67,577 clips from 2,463 unique speakers across 120 indigenous accents from 13 countries for clinical and general domain ASR, a benchmark test set, with publicly available pre-trained models with SOTA performance on the AfriSpeech benchmark.


Introduction
The African continent and the nearby islands constitute one-fourth of the land surface of the earth (Lodhi, 1993).Approximately 1.3 billion people live in Africa, which is about 18% of the world's population (Wikipedia contributors, 2023a).Of the estimated 7,000+ languages and dialects in the world, over 3,000 languages are found in Africa (Wikipedia contributors, 2023b; Heine and Nurse, 2000).
Despite its large and predominantly young population, Africa bears a significant proportion of the global disease burden (de Graft Aikins et al., 2010) with multiple socioeconomic factors contributing to high mortality and morbidity rates (Baingana and Bos, 2006).Healthcare systems are overburdened and underfunded in many African countries (Oleribe et al., 2019;Naicker et al., 2009;Nkomazana et al., 2015), struggling to cope with the increasing demand for services, while at the same time facing significant shortages of trained health workers (who; Ahmat et al., 2022;Naicker et al., 2010;Nkomazana et al., 2015;Kinfu et al., 2009;Etori et al., 2023).A recent study conducted by Ahmat et al. (2022) in 47 African countries shows that the region has a ratio of 1.55 health workers (physicians, nurses, and midwives) per 1000 people 3x less than the WHO-recommended density of 4.45 health workers per 1000 people.
While technology can help mitigate some of these problems, Bukachi and Pakenham-Walsh (2007) and Manyati and Mutsau (2021) aptly show that although Africa has enjoyed massive growth in mobile technology, telecommunication, and internet penetration over the past two decades, healthcare technology lags significantly.
A 2019 systematic review on the use of Automatic Speech Recognition (ASR) for clinical documentation in the US from 1990 to 2018 by Blackley et al. (2019) and other similar studies (Goss et al., 2019;Blackley et al., 2020;Ahlgrim et al., 2016;Vogel et al., 2015) showed that the use of speech recognition led to a 19-92% decrease in mean documentation time, 50.3-100% decrease in turnaround time, and 17% improvement in documentation quality.However, in the African context, the lack of training datasets for many of the 3000+ languages and accents in the continent remains an arXiv:2310.00274v1[cs.CL] 30 Sep 2023 obstacle in developing and adopting robust speech recognition systems for the general domain and for clinical ASR in particular (Doumbouya et al., 2021;Siminyu et al., 2021;Babirye et al., 2022;Ogayo et al., 2022).While recent efforts have begun to turn this tide for the majority of African languages like Swahili, Kinyarwanda, and Yoruba (Gutkin et al., 2020;Dossou and Emezue, 2021;Olaleye et al., 2022), over a thousand African languages and accents remain excluded from global speech research advancements.
Recent single-digit word error rates (WER) (Chen et al., 2022;Radford et al., 2022;Hsu et al., 2021a;Baevski et al., 2020b) in multiple SOTA publications and benchmarks on Librispeech (Panayotov et al., 2015a), TED-LIUM3 (Hernandez et al., 2018), and other datasets using architectures like Wav2vec2 (Baevski et al., 2020b), Conformer (Gulati et al., 2020), Transducer, and Whisper (Radford et al., 2022) contrast significantly with ASR performance for African accented speech (Gutkin et al., 2020;Dossou and Emezue, 2021) (see Figure 2).We explore whether curating a large pan-African speech corpus might unlock comparable single-digit performance on African accents.We restrict this investigation to accented speech in English because English is the official language for the medical record in most Anglophone African countries, expanding the utility of this work to multiple Anglophone African countries.
Our contributions are as follows: • We present AfriSpeech-200 1 , the first and most diverse open-source pan-African accented English speech corpus for clinical and general domain ASR, providing 200.70 hrs of accented speech, 67,577 speech-transcript pairs in 120 African accents across 13 countries, a benchmark dataset that paves the way for out-ofdistribution, few-shot and zero-shot analyses on very-low-resource accents.• We present a templating framework to augment existing corpora with native African proper nouns and evaluate multiple SOTA pretrained models and leading commercial ASR systems on our benchmark dataset.We provide in-depth analysis of selected models to 1 https://huggingface.co/datasets/tobiolatunji/afrispeech-200 2 AfriSpeech-200 is licensed under a CC BY-NC-SA 4.0 license explain their failure modes and offer helpful insights.
• We fine-tune the best-performing open-source models and achieve SOTA performance on the AfriSpeech benchmark dataset (108 African accents) as well as show promising zero-shot performance on very low-resource accents.
We provide best models3 as publicly available pre-trained checkpoints.
To enhance the performance of accented speech recognition, various methods have been proposed, which can be categorized into modeling and dataset approaches.On the modeling front, there have been efforts such as dialect-aware ASR models (Yadavalli et al., 2022), domain adversarial training (DAT) (Sun et al., 2018), combining DAT with transfer learning (Chen et al., 2020), using voice conversion (VC) (Zhang et al., 2022), combining VC with speed perturbation (Zhang et al., 2022), and accent pre-training (Acc-PT) (Das et al., 2021).These efforts, however, produced marginal improvements and still exhibit poor generalization capabilities.
Datasets have played a major role in improving ASR performance.The current SOTA in ASR (Radford et al., 2022) demonstrated the superior utility of large supervised datasets.Therefore, to bridge the ASR performance gap for African accented speech, multiple dataset creation efforts (Doumbouya et al., 2021;Siminyu et al., 2021;Babirye et al., 2022;Ogayo et al., 2022;Gutkin et al., 2020;Dossou and Emezue, 2021;Afonja et al., 2021;Kamper and Niesler, 2011;Ibejih et al., 2022) have been established.However, many of these datasets are limited in size and diversity.For example, Common Voice (Ardila et al., 2020)  While previous works have primarily focused on adapting Western accents to African accents, to the best of our knowledge, there has been limited research specifically addressing domain adaptation from a general domain to the clinical domain in the African context.In this regard, our work is the first attempt to bridge this gap and tackle the unique challenges associated with adapting accented African English ASR systems to the clinical domain.

AfriSpeech Dataset
We introduce AfriSpeech, a Pan-African accented English speech dataset for clinical and general domain ASR crowd-sourced from 2,463 African speakers, 200.70 hrs with an average audio duration of 10.7 seconds.Speaker, gender, age group, and clip domain distributions are shown in Table 2.In the following subsections, we describe the dataset creation process.

Focus Languages
We conducted an investigation on 120 African accents across 13 countries including the United States and Turkey.These accents originate from languages that belong to five language families, Since some languages are spoken across several countries (e.g., Swahili, isiZulu, Hausa, and Luganda), accents are not unique to countries.

Obtaining AfriSpeech Transcripts
Neural network models learn concepts from training data.Where the training data is predominantly Western (e.g.Common Voice (Ardila et al., 2019)), the resulting ASR systems fail to capture important pan-African contexts.For example, ASR systems fail woefully at transcribing African names like "Ogochukwu" (Igbo), "Malaika" (Swahili), or "Uwimana" (Rwandan), while excellently transcribing Western names like "Lauren" and "Bryan"representative of the bias in their training corpora.
To solve the problem of scarce African-centric text in the general and clinical domains, we created AfriSpeech using the following strategies.

Finding Available Transcripts
Our first task was to supplement existing large multi-domain corpora with African-centric text.
Our first target was Wikitext-103 (Merity et al., 2016), a collection of over 100 million tokens extracted from the set of verified "good" and "featured" articles on Wikipedia curated by Salesforce.
We split this corpus on sentence boundaries and randomly sampled sentences for our transcript corpus.Our next strategy was web scraping.We crawled and scraped major African news websites across multiple African countries on topics like politics, entertainment, sports, religion, education, etc.
In contrast to Wiki-text, the resulting corpus contained several African names, cities, and highly relevant vocabulary applicable to real-world use cases for downstream ASR.By scraping health-focused websites and health sections of news websites, we were able to get content from the clinical domain, albeit very little.
To increase clinical content representation, we focused on two multi-specialty biomedical datasets: PubMed (Wheeler et al., 2007) and NCBI disease (Dogan et al., 2014).We split these corpora on sentence boundaries and randomly sampled sentences for our transcript corpus.

Finding African Entities
We sourced for African-centric entities in two places: first, we leveraged an existing database of over 90,000 African names from the transatlantic slave trade between 1808 and 1863 (Anderson et al., 2013), which increased our coverage of African names, phonemes, and morphemes.We then used Okagbue et al. ( 2017)'s dataset of 965 Igbo names collected to reflect the dialectal classification of Igbo people and supplemented it with 1,000 more Nigerian names from other cultures such as Yoruba, Hausa, Fulani, Tiv, Efik, Ibibio, etc.These names were obtained from freely available textbooks, online baby name websites, oral interviews, published articles, and online forums like Instagram and Twitter.Finally, we obtained a list of African cities from Wikipedia (Wikipedia contributors, 2023c).

AfriSpeech Templates
The web scraping corpus was highly relevant but small.In the larger biomedical and Wikitext datasets, African content was sparse.We, therefore, sought to increase the utility of the curated corpora by creating "Africanized" versions.Several studies have demonstrated the utility of "templates" as an effective way to create richer, more expressive training datasets, especially for Question-Answering and prompt engineering (Pawar and Shrawankar, 2016;Brown et al., 2020;Yao et al., 2022) and named entity recognition (Davody et al., 2022).Inspired by this approach, we augment our dataset by sampling sentences from the corpora described above in addition to template sentences contributed by professional clinicians, hand-crafting a total of 140 template sentences.For each template sentence, we masked proper nouns (first names, last names, organizations, and cities), replacing them with their corresponding NER tags [PER, ORG, LOC].We then randomly replaced the masked tokens with African-centric entities-African names and cities, derived from section 3.2.2above, as well as common tropical diseases and medications.Each template sentence was reused 200 times.A random subset was sampled, sent as prompts for recording, and included with this release.Templated sentences represent approximately 30% of this corpus.

Audio Recording
Collection: Inspired by Common-Voice (Ardila et al., 2019) and SautiDB (Afonja et al., 2021), we developed and deployed a web-based application in Python/Flask (Figure 1) to collect crowd-sourced speech samples.The application also facilitates tracking of completion status, user demographics, reviews, and quality control.The app presents randomly selected sentences (prompts) to the speakers and prompts them to record their voices while reading the text.The speech recordings are persisted as mono-channel, 16-bit wav files, with a 48 kHz sampling rate.Post-processing tasks were performed on the audio recordings to remove samples shorter than 2 seconds and longer than 17 sec-onds.Raw unedited samples are provided as part of this release.Speakers in this dataset have been de-identified.Demographic information available includes gender, age group, accent, and country.
Annotation Instructions Recorder demographics are presented in Table 2. Instructions were provided to crowd-sourced recorders as detailed in Appendix A.2. Notably, the recorders were instructed to read punctuation marks in full and encouraged to use their natural accent.

Quality Control
Projects: Transcripts were bucketed into projects to separate clinical from general domain prompts.This approach maximized the time value of clinician contributors focusing their efforts more on medical prompts.

Reviewers:
We hired a team of human reviewers who up-voted or down-voted clips to indicate quality.Text feedback was also provided to recorders in 30% of cases where negative feedback was indicated.The text feedback contained the reason for the down-vote and was intended to help recorders improve future recording quality.
Guest Clip Review: New recorders were admitted as guests and allowed to record a maximum of 200 clips before quality review.10 to 30 clips were reviewed per guest and those who passed review were promoted to a "Paid" status.
Paid Clip Review: In the paid category, users were allowed a maximum of 200 clips before a temporary pause for quality check.During the temporary suspension, reviewers randomly reviewed 10% of the speech samples provided and positive, negative, or text feedback was provided.Access was restored if quality remained satisfactory, or users were blacklisted if over 30% of clips reviewed were down-voted.
Delisting Problematic Sentences: Where an audio clip receives a down-vote, the corresponding sentence is released for re-recording by a different user.If a clip recorded for the same sentence receives a second down-vote, the transcript itself is blacklisted.

Data
AfriSpeech-200 is a manually reviewed and curated subset, representing 7% of the total AfriSpeech As shown in Table 3, the train, test, and development sets are bucketed such that any given speaker may appear in only one.This ensures that contributors seen at train time are not seen at test time, which would skew the results.

Fine-tuning
Based on the benchmark results in For each model, we fine-tuned with FP16, AdamW (Loshchilov and Hutter, 2017), batch size of 16, for 10 epochs, with a linear learning rate decay to zero after a warmup over the first 10% of iterations.We fine-tune and evaluate on 3 domains: (1) general (25,812 clips), ( 2) clinical (41,765 clips), and (3) both (67,577 clips).We train on each domain and test across all 3 domains to investigate the effect of out-of-domain data on model performance.XLSR models were trained on a single Tesla T4 GPU with 16GB GPU memory while Whisper and Conformer models were trained on RTX8000 GPU with 48GB GPU memory.Fine-tuning took 24-48 hrs for all domains.

Model Vocabulary
Most pre-trained models define a limited vocabulary of only Latin alphabets with no numbers or punctuations (Baevski et al., 2020b).In stark contrast, numbers are critical in healthcare, e.g.blood pressure 130/80mmHg, or Lab results 0.428 mmol/L.Eliminating all numerical references in clinical text is dangerous and counterproductive.Post-processing to convert all numerical values to long form is imperfect so we retain numbers in their original form.For fine-tuning experiments, we define an alphanumeric vocabulary with semantically important punctuations, characters, and symbols commonly used in medical practice (colon, question mark, plus, etc).

Evaluation
We report our results as WER on AfriSpeech dev and test sets in addition to domain and accentspecific performance.Results are compared with Librispeech (Panayotov et al., 2015a) test set performance.We also report the zero-shot performance of fine-tuned models on unseen accents in the test set.

Africa-centric Fine-tuning Improves Robustness
As shown in

Training Data Bias
In the Open-Source section of Table 4, AfriSpeech dev and test set performance correlates with the number and diversity of pre-training datasets.For example, Wav2vec2 models trained exclusively on Librispeech significantly underperform when compared with those trained on multiple (Baevski et al., 2020b) or multilingual corpora (Babu et al., 2022).Models trained on Multilingual or multi-task corpora (Radford et al., 2022;Gulati et al., 2020) learn more useful representations, are more linguistically diverse, are more robust, and generalize better to accented speech.

Clinical ASR is Sensitive to Model Vocabulary
As mentioned in Section 4.4, most ASR models tend to transcribe numbers in their extended forms, which have a detrimental effect on their WER as shown in  transcribing numbers in clinical speech and converting them into correct numeric representations.

Punctuation Prediction is Critical for Clinically Useful ASR
Medical documents typically follow preset sequence and formatting, for example, patient history, general examination, laboratory investigation, etc., separated by new lines, section titles, or semicolons.Punctuation commands such as "Next line", "full stop" (.), "query" (?), "comma" (,), "colon" (:) are frequently used in healthcare dictations to add structure to documents.ASR systems without support for such commands force clinicians to review every line of the ASR transcript to add/revise punctuations and document structure, prolonging documentation time and patient wait time (Sunkara et al., 2020).As a result, commercial clinical ASR systems supporting these commands are preferable and outperform general-purpose models.

Commercial ASR APIs are Not So Global
The 3 large commercial ASR systems evaluated in this study have global presence.Millions of African Android users have access to Voice typing through the Google keyboard and Microsoft Word users have access to its ASR engine.on majority African accents and we show that despite their global presence, performance lags significantly on some of Africa's most populous accents like Swahili and Yoruba.

Domain Adaptation
Pre-trained whisper models performed better on general domain speech (AfriSpeech-general) when compared with the clinical domain, demonstrating the relative domain-driven difference in difficulty despite the robust training data for Whisper models (680k hours, 90 languages).Cross-domain finetuning yields significant gains helping to somewhat bridge this gap.Our results agree with prior work on domain adaptation (Sun et al., 2017;Abdelwahab and Busso, 2015) showing that models trained exclusively on clinical data improve when general domain data is added.Whisper shows 9% relative improvement on the clinical domain with the addition of general domain data.However, this trend is reversed with general domain data.Adding speech from the clinical domain leads to a 3% and 18.2% relative drop for Whisper and xlsr-53 respectively.Domain adaptation is no silver bullet.Care must be taken to apply this approach where benefits outweigh risks.

Accent-level Performance
Table 6 shows test set performance on the top 23 AfriSpeech accents grouped by their language families.We report the results for open-source, commercial, and fine-tuned ASR models.Fine-tuned models (ours) average relative improvement is 26.7% over the open-source ASR models and 36.5% over the commercial ASR models.For several accents, we observe that the whisper model fine-tuned with our AfriSpeech dataset shows the best overall performance with an average relative improvement of 16.2% across all accents, except in 4 South African languages (Zulu, isiZulu6 , Tswana, Afrikaans), Luo, and Kinyarwanda, where the fine-tuned model under-performs compared to the pretrained whisper model and commercial Azure model performs best on Luo accent.Although counter-intuitive, it is possible these accents are highly represented in Whisper pre-training data and require further investigation.

Zero-Shot Performance
We further explore generalizability to unseen accents, i.e., out-of-distribution (OOD) accents.Table 5 shows the results for the top 20 OOD accents in the test set.We observe an impressive 44.4% relative performance improvement across all OOD accents with our fine-tuned Whisper model compared to the baselines and 49.8% average relative improvement over the commercial models (Azure, GCP, AWS).These results demonstrate significant generalizability gains are achievable with better training data diversity.
5.9 Take SOTA LibriSpeech Results with a Grain of Salt North-African Accents are not included in this work.Because of the distinct nature of those accents, performance on sub-Saharan accents may not necessarily generalize to the Northern African Region.
Self-reported Accents: Similar to Common-Voice, recorders self-report their native tongue in free-text making it difficult to map to ISO-3 in all cases.Some users also reported their accents as "French", "English", "South African English", or a combination of accents.Although we attempted to clean and normalize the self-reported languages, this process was by no means perfect.As a result, accent names sometimes overlap e.g.Zulu and IsiZulu.Further cleanup could be done to consolidate these closely related accents.The dataset release will therefore include a normalized accent field for each sample.
Medical Abbreviations are Inconsistent: Since crowd-sourced recorders had varying levels of familiarity with the prompts, abbreviations like "Breast CA" may be pronounced fully as "Breast Cancer" or "Breast see-A".Since abbreviations abound in medical text and WER is not robust to such idiosyncrasies, models with correct predic-tions, e.g."Breast Cancer" are sometimes wrongly penalized where the transcript reads "Breast CA".
Integrating ASR in Healthcare Settings is Challenging: Cloud-based ASR presents some wellknown challenges in healthcare.Privacy is a major concern as there is a risk of unauthorized or malicious third-party access to confidential patient information.Furthermore, the perceived higher value of healthcare data among malefactors also heightens security risks for hospitals and ASR vendors.Additionally, Unethical ASR vendors could misuse confidential data for model training and development without proper consent.

Ethical Considerations
Clinical ASR models can improve productivity for clinicians, they can also increase documentation errors, especially through incorrect transcription of numbers, fractions, dates, and proper nouns which have legal, safety, and prognostic implications in healthcare.We caution clinicians to use ASR with full discretion and review transcripts carefully before final submission into the medical record.We release AfriSpeech hoping that it will be beneficial to clinical and non-clinical use cases within and outside Africa, improving ASR performance for accented speech and it may contain biases due to publicly available datasets.We do not have access to reviewers who are native speakers of most of the languages covered in AfriSpeech who can provide a rigorous review of self-reported accents.This hinders our ability to investigate samples from all languages.We hope that future users of the dataset will further investigate AfriSpeech's utility and quality for their languages."10th Dec, 1999", "10-12-1999", "Mon 10 Dec, 1999", "Monday 10th December, 1999".Similar timestamp variations were added to our templates.
Cleaning: Final corpus was pre-processed and cleaned by splitting on sentence boundaries, normalizing spaces, removing carriage return characters, removing non-alphanumeric characters except those with important structural or semantic meaning in the clinical domain such as question marks, parenthesis, colon, a hyphen, plus sign, and greater/lesser than sign.We removed Transcripts with less than 5 characters and greater than 300 characters.
Privacy and Patient information: Although the clinical corpora used were already anonymized, we re-examined several sentence samples for inadvertent exposure of patient names.Anonymized datasets with de-identification tokens like [NAME] and [DATE] were replaced with African names and randomly generated dates as described above.

A.2 Annotation Instructions
Recorders were provided with the following in-Accuracy It is very important that the recorded words match the text in the script exactly.If you accidentally deviate from the script, become unsure, or lose track of your thought, please delete and record the prompt again.
Punctuations All punctuations should be pronounced in full, not just observed.That is, when reading a text sample that contains punctuation, you say "comma", "full stop", "semi-colon", "colon", "slash", "hyphen", "question mark", "exclamation mark", and so on as appropriate.Brackets should be pronounced as "open bracket" or "close bracket".
Punctuation Exclusions/Exceptions to the above rule: In measurements or units like "mg/dl", please say "milligram PER dl" NOT "milligram slash dl".In situations where "?" is used to represent "query", please say "query" NOT "question mark".
Abbreviations Pronounce common short-hand forms (such as r/o, prn, tds, PO, mg, W/O), dates, times, and numbers as you would in a clinical setting.For example, "r/o" should be pronounced "rule out" as usual not "arr slash ohh".Common Abbreviations SHOULD be pronounced in full."CT" should be pronounced "see tee" as usual NOT "Computed Tomography"."CXR" should be pronounced "Chest Xray" as usual NOT "see ex arr" "mmHg" should be pronounced in full as "millimeters of mercury".Pronounce CA as "Carcinoma" NOT "See Ay".
Tone Also be sure to use your natural accent.The goal is to build a speech-to-text system that understands African accents.This tool is for us.Be natural.
Speed Do not speak unrealistically fast.While an increased reading speed is recommended, take care to avoid vocal fatigue from rushing through the phrases at lightning speed!This will only result in a lower-quality voice.Record a maximum of 2 hours a day, taking a break every half hour.

A.3 Annotator Management
Consent Recorders signed a Terms of Use agreement and consented to the privacy policy on the recording platform.
Payment Recorders were paid $5 to $10 per hour depending on task difficulty and clinical experience.
Most recorders considered payment satisfactory compared with task difficulty.

Figure 2 :
Figure 2: WER on LibriSpeech vs AfriSpeech for selected pre-trained models and commercial ASR systems.

Table 3 :
Dataset splits showing speakers, number of clips, and speech duration in Train/Dev/Test splits.
dataset, intended as an initial public release to stimulate research into African clinical and general domain ASR for accents with little or no representation in speech research.Table1shows the distribution of clips, unique speakers, and hours by country.

Table 4 :
Results showing selected models, number of parameters, Number of pre-training/fine-tuning corpora ["Multi" refers to multilingual or multi-task], Librispeech (Panayotov et al., 2015a) test clean WER and AfriSpeech dev and test set performance for open-source, commercial ASR models, and fine-tuned models (Ours).Missing values indicate incomplete or failed experiments.

Table 6 :
Table 6 compares the performance of these ASR APIs Test set performance per accent for open-source, commercial, and fine-tuned ASR models.