Biniyam

Building voice AI for African languages @addis_ai

[Biniyam @B1n1yam]

Addis Voices 2 now speaks seven African languages.

Swahili, Hausa, Wolaytta, Sidama, Amharic, Afaan Oromoo, and Tigrinya, across 56 voices.

Built for real-time voice agents, AI workflows, apps, educational content, audiobooks, and more.

Available through our Playground, API, and SDKs.

[Sam Ivere @hsprafrique]

I built ORIKA without Opus 5.5 or GPT Astra, while Grok 4.7 took over 10mins and I had to abandon it.

Guess the model used here

[Biniyam @B1n1yam]

@shalom4evermore Amharic speaker ነኝ btw, lmk where it went wrong

[Biniyam @B1n1yam]

Addis Voices 2 now speaks seven African languages.

Swahili, Hausa, Wolaytta, Sidama, Amharic, Afaan Oromoo, and Tigrinya, across 56 voices.

Built for real-time voice agents, AI workflows, apps, educational content, audiobooks, and more.

Available through our Playground, API, and SDKs.

[Biniyam @B1n1yam]

@ronin_skin16718 Thanks, Man! The voice models aren't open source, but they're available through our API and SDKs.

Docs: https://t.co/xRV3OGBCKf

[Biniyam @B1n1yam]

@unboundscripts Somali is already in training alongside several other low-resource languages. We're focused on getting the quality right before releasing anything. We hear you, and we appreciate the interest!

[Biniyam @B1n1yam]

@gurmessal That's a fair concern. Our Terms prohibit voice impersonation, deceptive AI-generated audio, and the use of our models to spread disinformation. We also require disclosure when AI-generated audio could mislead people. Reported misuse can lead to account suspension.

[Biniyam @B1n1yam]

@Kamaldacas We hear you, bro! Somali hasn't been forgotten. We'll share more when we're ready.

[Biniyam @B1n1yam]

@engineerdagi Thank you Dagi 🙏 You can hear all seven languages live here: https://t.co/9IRq9JIyXe

[Biniyam @B1n1yam]

@ethioel Absolutely😄 💯

[Biniyam @B1n1yam]

@africa1896 It's not opensouce but you can try it live https://t.co/9IRq9JI17G

[Biniyam @B1n1yam]

https://t.co/9IRq9JI17G you can try it here

[Biniyam @B1n1yam]

Weights and benchmark code are on HF, so you can run the numbers yourself.

https://t.co/KP9EWTn8mQ

[Biniyam @B1n1yam]

We’re open-sourcing Addis Scribe for Afaan Oromo.

It beats Google Chirp and Gemini Argon on accuracy.

[Biniyam @B1n1yam]

It’s also available through our SDK and API.

https://t.co/zFz1eNGhBr

[Biniyam @B1n1yam]

Weights and benchmark code are on HF, so you can run the numbers yourself.

https://t.co/KP9EWTn8mQ

[Biniyam @B1n1yam]

It’s also available through our SDK and API.
https://t.co/zFz1eNGhBr

[Dagmawi Babi @DagmawiBabi]

Introducing DMs in @ScholarXIV
• https://t.co/rIFfgW9Pw2

This's the first of many collaborative research features we're building. Search by usernames or emails, and starting exchanging messages, sharing files, sending papers and collections, collaborating on your research and much more.

You can also create groups and add your team to work on your research together. Groups can also have topics (like Slack) to help you categorize your conversations.

A unique feature we've added is an agent, accessible in every DM or Group. We call it @Scholar and it's your personal research agent brought within chats. @Scholar knows the history of the chat and who's talking to it, and accordingly will switch it's context and memory based on your personal chats with it.

In any chat just tag @Scholar and it will respond to you. Consider @Scholar as your peer researcher to help you and your team progress in your research.

[Biniyam @B1n1yam]

@claudeai We got accepted into Claude Startups.

For a team building voice infrastructure for low-resource African languages, a year of Claude Team, API credits, and office hours with Anthropic’s Applied AI team is real runway.

Thanks @claudeai

[Biniyam @B1n1yam]

@chapimenge3 That Hohe inference pipeline you dropped definitely helped bro 🙌

[Biniyam @B1n1yam]

We open-sourced Addis Scribe Streaming, realtime speech recognition for Amharic.

It beats Google Chirp 3 on both accuracy and speed:

Word error rate: 28.4% vs 32.5%
First words on screen: 1.56s vs 6.67s
Final text after you stop talking: 30ms vs 1,411ms
Open weights, and available through our API and playgrounds.