Introducing the Perigee-1 Family
Today, we are introducing the Perigee-1 family of models: a new generation of speech and language AI from OrbitalsAI, built African-first and designed for production.
Perigee started with a simple problem: speech and language infrastructure still works poorly for many of the people and languages it is supposed to serve.
Over time, that problem expanded beyond transcription. Developers need real-time speech recognition, translation, natural voice generation, structured understanding, privacy controls, and language models that can work directly across African languages without treating them as an edge case.
Perigee-1 is our answer to that.
It is a family of four models covering transcription, real-time speech, text-to-speech, and multilingual text intelligence — available today through the OrbitalsAI API, Playground, and Live Studio.
African language AI should not be an edge case
Most language infrastructure is still designed around a small group of dominant languages and usage patterns.
That assumption breaks quickly in Africa.
Conversations move between languages. Accents vary by region. English, Hausa, Yoruba, Igbo, Swahili, Arabic, Pidgin, and other languages can appear in the same interaction. The way people actually speak is often more complex than the clean, single-language inputs that many systems expect.
We think the infrastructure should adapt to people, not the other way around.
Language infrastructure should adapt to people, not the other way around.
Meet the Perigee-1 family
Perigee-1 is not a single model. It is a family of models designed to cover the speech and language workflows developers need in production.
The family starts with speech recognition, but extends across real-time transcription, voice generation, translation, summarization, structured extraction, privacy, and conversational language tasks.
The four models are:
Perigee-1-transcribe for recorded speech recognition.
Perigee-1-transcribe-realtime for low-latency streaming speech recognition and translation.
Perigee-1-tts for multilingual speech generation and zero-shot voice cloning.
Perigee-1-text for translation, summarization, structured extraction, PII redaction, and chat.
Perigee-1-transcribe
Perigee-1-transcribe turns recorded speech into structured text across African and global languages.
It accepts common audio formats including MP3, WAV, M4A, FLAC, AAC, and OGG, with files up to 200 MB. Responses can include the transcript, word-level timestamps, and subtitle-ready SRT output.
The model is built for the conditions speech systems actually encounter: regional accents, noisy recordings, varied speaking styles, and conversations that do not stay neatly inside one language.
This matters especially in markets where code-switching is normal rather than exceptional. A conversation may move between English and Hausa, Yoruba and English, or other language combinations without warning. The transcription system has to follow the speaker rather than force the speaker into a rigid input format.
Perigee-1-transcribe-realtime
For live applications, Perigee-1-transcribe-realtime brings the same speech recognition capabilities into a streaming environment.
Audio is sent over WebSockets and the model returns partial and final transcription results as speech arrives. End-to-end latency is approximately 200 ms, making it suitable for interactive applications where waiting for an entire recording to finish is not an option.
Realtime can also translate speech as it is being transcribed. This makes it possible to build multilingual meetings, customer-support systems, live captions, voice interfaces, and other applications where language needs to move with the conversation.
Perigee-1-tts
Perigee-1-tts turns text into natural speech across 19 supported languages.
It generates 24 kHz, 16-bit WAV audio and supports zero-shot voice cloning, allowing developers to create new speech in a target voice from a short reference sample.
The goal is not only to produce understandable speech, but to make voice interfaces feel usable across languages that are still poorly served by mainstream speech infrastructure.
This opens up applications across accessibility, education, customer support, media, voice assistants, and any product where text needs to become speech naturally.
Perigee-1-text
Perigee-1-text handles the language tasks that happen after speech becomes text.
It supports translation, summarization, structured extraction, PII redaction, and conversational chat across the same multilingual system.
Translation is especially important. Perigee-1-text supports 19 languages and 342 direct source-to-target translation directions, without requiring English to sit in the middle of every language pair.
That means a Hausa sentence can be translated directly into Yoruba. It does not need to become English first and then be translated again.
Removing that forced pivot simplifies the pipeline and helps preserve meaning across languages that developers increasingly need to connect directly.
Translation is only one part of the model. Perigee-1-text can also turn unstructured language into structured outputs, summarize long conversations, remove sensitive information, and power multilingual conversational systems.
More than translation
Production language systems rarely stop at transcription or translation.
A customer-support call may need to be transcribed, translated, summarized, checked for sensitive information, converted into structured fields, and then passed into another application.
Perigee-1-text is designed to support those downstream tasks directly.
Structured extraction. Convert natural language into predictable JSON or other structured formats that software can consume.
Summarization. Reduce long conversations, transcripts, or documents into concise outputs while preserving the important information.
PII redaction. Detect and remove sensitive personal information before text moves further through a workflow.
Chat. Build multilingual conversational experiences across supported languages using the same API surface.
How Perigee-1 performs
We evaluate Perigee-1 across both Nigerian-language benchmarks and a broader set of African-language datasets.
Word error rate, or WER, measures the difference between a model’s transcription and the reference transcript. Lower is better.
On NaijaVoices, Perigee-1 performs strongly across Hausa, Igbo, and Yoruba, including against larger general-purpose speech systems.
The results are not uniform across every model and every language, and that is important.
Perigee-1 leads the comparison on Hausa and Igbo in this evaluation. On Yoruba, Sahara performs slightly better. We would rather show that difference than hide it.
Our goal is not to claim that one model wins every benchmark. It is to build speech infrastructure that is consistently strong across the languages and conditions our users actually encounter.
We also evaluated Perigee-1 across additional languages using established datasets for Kinyarwanda, Twi, Swahili, and Arabic.
These results show the same pattern. Perigee-1 is competitive across a wider range of African languages, but it does not lead every benchmark.
On Kinyarwanda, Perigee-1 records a lower WER than Sahara. On Twi, Sahara performs better. On Swahili, ElevenLabs records the lower score in the comparison, and on Arabic, QuartzNet performs better on MediaSpeech.
Different datasets have different recording conditions, speakers, domains, and difficulty levels, so scores across languages should not be treated as directly comparable. What matters to us is building a system that performs reliably across a broad and growing language set rather than optimizing for a single leaderboard.
Designed to work as a system
The individual Perigee-1 models can be used independently, but they are designed to compose.
A speech application can begin with audio, transcribe it, translate it, remove sensitive information, extract meaning, generate a response, and return that response as speech — without stitching together unrelated language systems from multiple providers.
Each step can also be removed, replaced, or used on its own. A developer building live captions may only need realtime transcription. A call-center workflow may need transcription, redaction, and summarization. A voice assistant may use the entire pipeline.
The point is not to force every application into one workflow. It is to make the underlying speech and language primitives available through one coherent family.
Built for developers
Perigee-1 is available through a consistent API surface designed for applications that need to move between speech, text, translation, and voice.
Developers can start with a single model and add more capabilities as the product grows. Recorded transcription, real-time speech, translation, text intelligence, and speech generation can all be used independently or composed into larger workflows.
The Playground provides a fast way to test models interactively, while the SDK and API are designed for production integration.
Python — Transcribe audio
The same SDK extends across the rest of the Perigee-1 family. Translation, streaming speech, and voice generation use the same authentication and developer workflow, so teams can add capabilities without rebuilding their language stack from scratch.
Python — Hausa to Yoruba translation
I would keep the realtime and TTS snippets in Contentful for future articles/docs, but not put all four into this launch post. Two examples are enough to establish that this is a real developer product without turning the announcement into documentation.
Available today
The Perigee-1 family is available today through the OrbitalsAI platform.
Developers can use the Playground to test recorded transcription, translation, text-to-speech, and other language capabilities interactively. Live Studio provides a real-time environment for testing streaming transcription and translation as speech happens.
For production applications, the same capabilities are available through the OrbitalsAI API and SDKs.
The current family includes:
Perigee-1-transcribe — recorded speech recognition Perigee-1-transcribe-realtime — streaming transcription and translation Perigee-1-tts — multilingual speech generation and voice cloning Perigee-1-text — translation, summarization, structured extraction, PII redaction, and chat
Together, the family supports 19 languages across speech and text workflows, with 342 translation directions available through one translation API.
Language infrastructure for more of the world
Speech and language AI is becoming infrastructure.
It will sit inside customer support systems, education products, financial services, healthcare tools, media platforms, accessibility software, developer tools, and interfaces we have not built yet.
The question is not whether these systems will become widespread. It is who they will work for when they do.
We built Perigee-1 because African languages should not be treated as an integration to add later. They should be part of the infrastructure from the beginning.
That means building models around the languages people speak, the accents they use, the way they switch between languages, and the conditions in which real conversations happen.
Perigee-1 is our first complete step toward that infrastructure.
Africa should not be added to that future afterwards. It should be part of the infrastructure from the beginning.
Start building with Perigee-1