Today, we’re excited to officially launch our Yoruba speech data campaign on Dialectra. Over the past two months, we’ve seen contributors across Hausa, Kanuri, and Fulfulde help us build one of the fastest-growing African speech data communities. Now it’s time to expand. With Yoruba joining Dialectra, contributors can now participate in: • Corpus script recordings • Transcription tasks • Live conversational speech through Dialect Connect As always, every contribution goes through our transcription, annotation, standardization, and human verification pipeline before becoming training-ready datasets. Yoruba is one of Africa’s most influential and widely spoken languages, yet high-quality conversational speech infrastructure for it remains limited. We want to help change that. If you speak Yoruba, you can now join Dialectra, contribute your voice, and help shape the future of African speech AI while earning rewards for your contributions.

Lately, I've seen more founders talking about solving African speech and language infrastructure. That's a good thing. The opportunity is far bigger than any one company. But I also think we're approaching it the wrong way if everyone tries to build in isolation. No single team can capture the diversity of African languages and dialects alone. It takes researchers, communities, contributors, universities, startups, and open collaboration. That's one of the reasons we started releasing portions of Dialectra's datasets as open source on Hugging Face. We want researchers and builders to have a foundation they can build on, while we continue developing larger datasets and infrastructure for enterprise use. Open collaboration doesn't weaken the ecosystem—it strengthens it. If we genuinely want African languages to be represented in the next generation of voice technology, we need to build more together and compete less on the fundamentals. We're still early, still learning, and we'll continue contributing where we can. huggingface.co/Dialectra

More speech data doesn't mean better speech AI. Let me give example with Hausa, one fo the largest datasets we have followed by Yoruba, Hausa isn't just another language to throw into a voice model and expect it to work. It has tones, implosives, and consonant clusters that most ASR systems have never seen, e.g ɓ → ɓera, ɓarna ɗ → ɗaki, ɗalibi ƙ → ƙasa, ƙofa ƴ → ƴanci, ƴatsa sh, ts, gw, gy, ƙw, kw, ky, ƙy... Every single one of these is phonetically distinct, miss one and your model breaks. So we don't just collect voice data. Every submission in our pipeline is tagged with: → Full transcription text → Gender & age group → Dialect region → Phonetic metadata Random data collection is not a dataset. STRUCTURE IS. All our datasets goes through human-in-the-loop verification, our annotators aren't just listening for audio quality, they're checking phonetic accuracy, transcription fidelity, and dialect consistency for every single submission. And we still don't stop there. After human review, every dataset runs through WER and CER measurement to catch anything the human pass might have missed. We believe two layers. No compromises. Because we're not building a data collection platform. We're building the infra layer for African voice AI. The foundation that didn't exist until now.

Today we're making part of our Yoruba speech datasets publicly available for the community. It's a small release compared to what we've collected so far, but we believe African language AI advances faster when researchers, developers, students, and startups have access to quality datasets. Over the last few months, Dialectra has grown to more than 500,000+ voice datasets samples and over 500+ hours of speech data across multiple collection workflows. One of the most exciting contributors to that growth has been Dialect Connect, our conversational speech platform that captures how people actually speak not just scripted recordings. We're committed to releasing datasets where it makes sense and where contributor permissions allow. For organizations, research teams, and enterprises looking for larger datasets, custom collections, dialect-specific corpora, conversational speech data, transcription projects, or benchmarking support, our team has a structured onboarding process and would be happy to discuss your requirements. We're not only collecting speech data we're building the infrastructure layer for African voice AI. HF:huggingface.co/datasets/Dial…

We released less than 0.05% of our Hausa speech dataset. Within days: researchers, startups, enterprises, and builders were in our inbox. That's not a coincidence. High-quality African speech data is one of the most valuable layers in AI right now and almost nobody is building it seriously. Dialectra is. Enterprise access: Dialectra.io And next up: our first Hausa TTS model, trained on Dialectra data, available via API. The real work is just beginning.

We're preparing to release our first Hausa TTS model, trained on Dialectra datasets and designed to give builders, developers, and Hausa content creators access to real Hausa voice AI through an API.

Today, Dialectra releases its first open Hausa voice dataset on Hugging Face for free. This is a small but meaningful step toward something bigger: dialect-aware AI infrastructure for African languages. Hausa is one of the most widely spoken languages in Africa, with over 70 million speakers across Nigeria, Niger, Ghana, and beyond. Yet most voice AI systems fail Hausa speakers not because the technology can't work, but because the right data doesn't exist. Dialectra is building that data layer. This first dataset includes: → Dialect-tagged audio recordings from native Hausa speakers → Rich metadata: region, speaker profile, speaking style, quality score → Benchmark-ready structure for ASR and TTS development We're making it available to researchers, AI teams, and developers who are building voice technology that actually works for African users. This is just the beginning. More languages, more dialects, more infrastructure coming soon. Dataset: huggingface.co/datasets/Dial… Learn more: Dialectra.io

We are moved to the next step as start up 🤝 Thank you for believing in us.

Slowly building the team behind Dialectra. Happy to welcome a new AI/ML Engineer as we continue laying the foundations for African speech AI. We've spent months building datasets, annotation workflows, benchmarking systems, and speech infrastructure. Now it's time to push deeper into models, evaluation, and product development. Welcome to the journey @nababa_h

Thanks Msageer for the clear and powerful Article ! It perfectly explains why we’re focused on building real dialect-aware data for African languages and beyond. If you want AI that actually understands every voice, jump in at dialectra.io and start contributing. Let’s build this together.

Are you familiar with our new update? Dialect Connect lets you invite a partner, discuss a topic together, and submit when you're done. Collaborate. Contribute. Earn.

One thing we’ve been testing heavily at Dialectra recently is dialect benchmarking across Hausa speech cohorts. A lot of people outside Northern Nigeria see “Hausa” as one language, but in reality Hausa contains multiple dialect variations with different pronunciations, vocabulary patterns, accents, and speech rhythms. But once you start working deeply with speech datasets, you quickly realize how different Kananci, Sakkwatanci, Katsinanci, and other dialect cohorts can sound in real conversations. Different pronunciations. Different rhythms. Different word usage. Different speech patterns. And the impact on model performance becomes very obvious. Recently we started running deeper internal benchmark evaluations across Hausa dialect cohorts inside our DIE Engine (Dialectra Intelligence Evaluation). Some cohorts performed surprisingly well. Others showed much higher WER than expected even with the same base model and similar datasets. That was a big reminder for us: building African speech infrastructure is not only about collecting more audio. It’s about understanding linguistic variation properly. This is also why we’ve been investing heavily into: * human review pipelines * dialect-aware annotation * normalization systems * conversational datasets * benchmark reporting Because a model that performs well on “generic Hausa” may still fail badly when exposed to real regional speech patterns. Still early, but these evaluations are helping us understand where the actual infrastructure gaps are. And honestly, this part of the work has become one of the most interesting things we’re building at Dialectra.

We're building AI that understands you in your language. Every recording you submit on Dialectra trains AI to recognize Hausa, Yoruba, Arabic, and more. You're not just a user. You're a builder. Join #Dialectra today: dialectra.io/auth?tab=signup…

A quick update on how things are moving at Dialectra. ​We’ve officially crossed a few massive milestones: ​1,166 total conversational calls ​904 completed sessions ​142.2 hours of live speech collected ​Active contributors spanning multiple African language communities ​We’re also hitting some scaling pains. Some users are currently dealing with dropped calls and connection issues on Dialect Connect. We're already under the hood fixing the infrastructure to smooth this out. ​The wild thing is, Dialect Connect started as a pure experiment to see how live conversation stacked up against traditional data transcription workflows. What happened next caught us completely off guard. People didn't just treat it as a data tool; they used it to genuinely connect with each other, have real conversations, and earn rewards naturally. It completely changed how we think about conversational AI infrastructure. ​Building this as a bootstrapped founder has meant a lot of late nights, technical headaches, and raw uncertainty. But we didn't build Dialectra to catch a few months of hype and fade away. ​We’re still early, still learning, and still fixing bugs daily. But the foundation is laid, we’re here to stay, and the real work begins now.

Got questions? Need help with tasks? Join the Dialectra Telegram community for fast support, updates, and discussions with fellow contributors. t.me/dialectralab