AI transcription and speech recognition, built into your software

Your users speak, your software acts. Three voice agents ready to plug into your software, real-time or deferred, with no length limit, delivering a result ready to validate: a completed report, meeting minutes with their tasks in the CRM, a formatted letter.

Available via API: REST to integrate them into your software, WebSocket for real time, MCP to plug them into your own agents.

Already tried transcription? Quality is won after it.

You may have already built your own solution, or compared several transcription engines. The text is there, but your users expect more: their domain terms spelled correctly, their instructions followed, a document ready to validate. That level of quality comes from what the agent does after transcription.

A transcription API gives you text. Agora gives you a result.

  1. 1

    Transcribe

    Multi-model: the best model for each language and use case.

  2. 2

    Contextualize

    Your domain vocabulary, your proper nouns, your product codes, your files.

  3. 3

    Correct and structure

    Several AI correction passes, tailored to your business, then extraction of the key information (names, dates, amounts, references) in the expected format.

  4. 4

    Act in the software

    The report fills itself in, tasks are created, the file is updated.

Multi-model

We continuously evaluate speech-to-text models. When a better-performing model comes out, your software benefits with no change to your integration.

Who speaks, and when

Diarization tells speakers apart, then the agent names them as soon as their name is spoken in the conversation. Essential as soon as several voices are involved.

The right words

Domain glossary and data from your software: names, references, dates and amounts are written exactly.

Hosted in France

GDPR by design, no external provider. Recordings are encrypted in transit and at rest, and processed in France, by a sovereign provider.

Human in the loop

The agent prepares, the human validates

Whatever the agent, the people who spoke find, in a review interface, what was transcribed and what was created in the software, with points needing attention highlighted. They check what matters, correct if needed, and keep the final say.

Real time or deferred, as needed. And no length limit.

Real time

Text appears as people speak

At the pace of speech: users watch their dictation being written. A channel can also stay open all day, to transcribe a room or a phone line. This capability is still rare: many solutions transcribe after the fact, and those that offer streaming often cut the session after a few minutes.

Deferred

Your recordings processed in a queue, corrected and enriched

Meetings, appointments, voice recorder files: recordings are queued and processed one after another, when the accuracy and richness of the result matter more than immediacy.

No length limit

A three-hour meeting, a morning of dictation

The agent processes the whole recording, with no splitting and no session to restart. Caps are common on the market: meetings limited to 30 minutes or 2 hours, dictations cut after 5 to 20 minutes, 5-minute streaming sessions.

Your capacity, in channels

You split your channels between real time and deferred

Your subscription sets a number of channels. Each one serves real time (a dictation, continuous listening) or deferred processing (a queue of recordings). You change the split as your usage evolves, and scale your subscription if needed.

No per-minute or per-token billing: your subscription sets a capacity, independent of your consumption.

Where your users already speak

On mobile

In your software's app, in the field, between two appointments.

On desktop

From your software, at the office or during a meeting.

By audio file

A recording from a voice recorder, a video conferencing tool or any other capture tool is sent to the agent.

An API designed for fast integration

Real-time WebSocket API

Bidirectional audio streaming, multi-session. Fits into any web or mobile application.

Automatic language detection

Users speak in their own language: the system detects it and switches models, with no configuration.

  • Français
  • English
  • Deutsch
  • Español
  • Italiano
  • Português
  • Nederlands
  • 日本語
  • 中文
  • 한국어
  • العربية
  • Polski
  • Türkçe
  • Русский

Admin console

The occupancy rate of your channels, real-time and deferred: you adjust the split yourself, based on actual usage.

On your harness or ours

Your harness

You have built your own agentic harness: the Speak agents plug into it as tools, via MCP.

Agora harness

The agents are hosted, supervised and updated by Agora. You integrate them into your software via REST API, or via WebSocket for real time.

Another use for voice?

Voice commands, calls and voice messages, captioning: build your own voice agent on the Agora platform, with the same building blocks.

Frequently asked questions about AI transcription

What is the difference between speech recognition, AI transcription and speech-to-text?

All three refer to the same technology: turning speech into text. "Speech recognition" more broadly covers systems that respond to voice, including commands, "AI transcription" focuses on the resulting text, and "speech-to-text" (or "voice-to-text") is the term developers use most. Agora covers all three uses: dictation, transcription of recordings and voice commands.

What does Agora add to a transcription API?

A transcription API turns audio into raw text. Agora agents deliver the result: transcription takes your domain vocabulary into account, an LLM corrects and structures the output, and the agent puts it to work in your software, such as filling in a report, creating tasks or updating a file.

Is there a Whisper API alternative hosted in France?

Yes. Agora is a sovereign alternative to the Whisper API: the platform is hosted in France, multi-model ASR, and your audio never goes to an external provider. We continuously evaluate speech-to-text models and select the one that fits your language and use case.

Can meeting minutes be generated automatically?

Yes, that is the job of the meeting minutes agent. The recording is transcribed, speakers are told apart and named as soon as their name is spoken, then the meeting is summarized topic by topic: decisions, and tasks created in your software, assigned and dated.

Real-time or deferred transcription: which should I choose?

Real time fits when users expect an immediate response: dictation, voice commands, live captions. Deferred processing fits long recordings (meetings, interviews, calls), when accuracy matters more than latency. The platform offers both: you split the channels of your subscription between the two modes to match your usage.

Is there a recording length limit?

No. A meeting lasting several hours or a long dictation session is processed in full, in real time as well as deferred, with no splitting and no session to restart.

How is transcription billed?

By subscription, with no per-minute or per-token billing. Your subscription sets a number of channels, individually or in packs (5, 10 channels…), which you freely split between real time and deferred: your cost stays the same from one month to the next, and you scale the subscription as your usage grows.

Which languages are supported?

The main European and Asian languages, including French, English, German, Spanish, Italian, Arabic, Japanese and Chinese. The language is detected automatically, with no setting on the user's side.

How are audio recordings protected?

Recordings are encrypted in transit and at rest, and processed in France, by a sovereign provider. The platform is GDPR compliant, with no external provider.

Send us a real case, we'll show you the result

A recording from your users, and a few questions about their context.

What do your users dictate?
In what conditions?
Where should the result go?
How are recordings captured?
Have you already tested a transcription solution?

Your details are only used to get back to you about this request. Terms of service

Request a demo