Behind each of these systems lies a critical component: **large speech datasets**.
These datasets allow machine learning systems to learn how languages sound in real-world environments. They capture pronunciation patterns, accents, dialectal differences, and the natural rhythm of speech. Over the past decade, advances in machine learning have dramatically improved speech recognition and synthesis systems. However, these improvements have largely benefited **high-resource languages** such as English, Mandarin, Spanish, and French. Many of the world’s languages remain absent from these technological advances. ### 1.2 The Global Speech Data Divide Modern speech recognition systems require enormous amounts of training data. State-of-the-art systems often rely on **thousands of hours of annotated speech recordings** in order to achieve high accuracy. Unfortunately, most languages do not have such resources. Of the roughly **7,000 languages spoken globally**, only a small number have the large, well-curated speech datasets necessary to build robust speech technologies. This disparity creates what researchers often refer to as the **speech data divide**. Languages with abundant data continue to benefit from improved AI tools, while languages with little or no digital data remain excluded from technological innovation. For speakers of these underrepresented languages, the consequences are significant: - voice assistants cannot understand them - automated services are unavailable in their language - speech-to-text systems perform poorly or not at all - digital platforms fail to recognize their linguistic identities As voice interfaces become a primary way people interact with technology, this divide risks **deepening global digital inequalities**. ### 1.3 African Languages and the Data Gap Africa is the most linguistically diverse continent in the world. It is estimated that **over 2,000 languages** are spoken across the continent. These languages belong to several major language families. Many African languages are widely spoken by millions of people, yet remain **extremely underrepresented in digital datasets**. Even widely spoken languages such as Yoruba, Hausa, Amharic, and Swahili often have far fewer digital speech resources compared to European or Asian languages. The situation is even more challenging for languages with fewer speakers. Languages spoken by **less than one million people** often have little to no available speech data, making it nearly impossible to build speech recognition systems for them. Several factors contribute to this gap: 1. **Limited funding for language resource development** 2. **Scarcity of technical infrastructure for data collection** 3. **Lack of standardized methodologies for collecting speech datasets** 4. **Low participation of local communities in AI development processes** 5. **Barriers to publishing open datasets** As a result, many African languages remain **digitally invisible**. ### 1.4 Why Speech Data Matters Speech datasets form the foundation for multiple technologies that can benefit communities across Africa. These technologies include: ##### a) Automatic Speech Recognition (ASR) ASR systems convert spoken language into written text. These systems enable: - voice search - transcription services - automated customer service - accessibility tools ##### b) Text-to-Speech (TTS) TTS systems convert text into natural-sounding speech. These technologies support: - audiobooks - assistive technologies - voice assistants - language learning applications ##### c) Speech Translation Speech translation systems allow spoken communication to be translated into other languages in real time. ##### d) Conversational AI Speech datasets also power chatbots and conversational agents that interact with users through voice. When speech technologies support local languages, they can improve access to services in key areas such as: - healthcare - agriculture - education - disaster response - financial inclusion - public information services However, without speech datasets in these languages, such technologies cannot be developed. ### 1.5 Community-Driven Speech Data Collection In recent years, a growing number of initiatives have begun addressing the speech data gap for African languages. One of the most promising approaches is **community-driven data collection**. Rather than relying solely on centralized research institutions, community-driven approaches engage **native speakers, local organizations, and grassroots networks** to collect speech recordings. Projects such as **AfriSpeech-200 **have demonstrated the power of community participation in building large-scale datasets. AfriSpeech collected speech recordings from speakers across multiple African countries to create a pan-African accented speech dataset used for speech recognition research. Similarly, the **African Voices Kenya** project gathered approximately 30**00 hours of speech data across five Kenya languages** through community engagement and ethical data collection practices. These initiatives show that **distributed community participation can generate large and diverse speech datasets**, even in resource-constrained environments. Community-based approaches offer several advantages: - they capture authentic speech patterns - they include diverse dialects and accents - they empower speakers to contribute to digital representation of their languages - they create shared ownership of language resources # Overview ### **Purpose of the Voice Playbook** This playbook provides **practical, ethical, and scalable guidance** for collecting **voice datasets for African languages**, particularly **low-resource languages with limited digital presence**. The playbook enables: - Community-led voice dataset creation - Standardized speech dataset methodologies - Ethical and consent-based recording - Scalable workflows for NLP training - Reusable processes for future language initiatives The playbook supports **Automatic Speech Recognition (ASR)**, **Text-to-Speech (TTS)**, and **multimodal language models**. African languages remain severely underrepresented in speech datasets, limiting their participation in modern AI systems. Many African languages lack large-scale audio datasets necessary for training speech technologies. > **This playbook lowers barriers by enabling grassroots communities, universities, NGOs, and language activists to collect high-quality speech data.** # MODULE 1 # How to Use this Playbook This playbook is for anyone who wants to collect voice data in an African language, whether the project is: - Small — 50–100 speakers - Medium — 500–1,000 speakers - Large — thousands of contributors - Academic - Community-led - Government-led - NGO-led - Commercial - Open-source - Designed for ASR/TTS - Intended for language documentation - Intended for AI research It is designed to answer a simple question: ***“If I want to collect voice data in my language, what exactly do I need to do?”*** The playbook takes you through the entire lifecycle: > ##### Plan → Design → Prepare → Mobilize → Recruit → Consent → Train → Record → Validate → Transcribe → QA → Pay → Document → Release → Sustain Voice collection is not simply a recording exercise. It is an operational system involving people, language expertise, community relationships, technology, quality control, ethics and data management. The AfriVoices-KE project organized these functions through language leads, resource persons, mobilizers, contributors, validators, transcribers, super-reviewers and administrators. # PART I — BEFORE YOU COLLECT ANYTHING ### 1. Define Why You Are Collecting Voice Data Before recruiting anyone, answer: #### 1.1 What will the data be used for? Examples: - Automatic Speech Recognition - Speech-to-text - Text-to-Speech - Speech translation - Voice assistants - Language documentation - Linguistic research - Benchmarking - Conversational AI - Digital government - Healthcare - Agriculture - EducationYour intended use determines what type of recordings you need.
#### 1.2 What kind of speech do you need? There are two basic approaches used in the reference project. **a) Scripted speech** A participant reads a prepared sentence. **b) Unscripted speech** A participant responds naturally to a question or prompt. The reference project deliberately combined the two, targeting approximately 75% unscripted and 25% scripted speech. #### Practical recommendation If your goal is a general-purpose speech dataset, consider collecting both.Type | What it gives you |
Scripted | Controlled sentences, predictable content |
Unscripted | Natural speech, dialects, spontaneous phrasing |
Both | Broader coverage |
Metric | Target |
Speakers | 1,000 |
Total hours | 500 |
Scripted | 125 hrs |
Unscripted | 375 hrs |
Regions | 10 |
Dialects | 4 |
Female | 50% |
Male | 50% |
Age groups | 5 |
**“We are collecting Language X.”**
Ask: - Which variety? - Which dialects? - Which geographic areas? - Which communities? - Is the language internally diverse? - Are some varieties more widely used? - Are some varieties endangered? - Which varieties are appropriate for the intended use? > Example : The AFN project explicitly considered dialect variation in languages including Kalenjin, Maasai, Somali, Dholuo and Kikuyu. #### Create a language profile Record: 1. Language: 2. Language family: 3. Country/countries: 4. Main regions: 5. Dialect groups: 6. Estimated speaker distribution: 7. Existing digital resources: 8. Existing speech datasets: 9. Native-speaker experts: 10. Relevant institutions: #### 4. Define Your Dialect Strategy This is one of the most important steps. A dataset can contain thousands of hours and still be poorly representative if most speakers come from one location or dialect. Create a dialect matrix:Dialect | Geography | Target speakers | Target hours |
Dialect A | Region 1 | 200 | 100 |
Dialect B | Region 2 | 200 | 100 |
Dialect C | Region 3 | 200 | 100 |
Dialect D | Region 4 | 200 | 100 |
Domain | Scripted | Unscripted | Target hours |
Everyday life | ✓ | ✓ | 50 |
Agriculture | ✓ | ✓ | 100 |
Healthcare | ✓ | ✓ | 50 |
Education | ✓ | ✓ | 50 |
Government | ✓ | ✓ | 50 |
Stories | ✓ | 50 | |
Customer care | ✓ | ✓ | 50 |
Other | ✓ | ✓ | 100 |
**“Imagine you are calling a customer-care agent because your mobile money transaction failed.”**