Voices and recognition
Where an agent's voice, its speech recognition and its AI model are chosen, which vendors your platform offers, and what each language can use.
Every agent has three choices that decide how a call sounds and how well it understands the caller:
- Voice: how the agent sounds (text to speech).
- Speech recognition: how the caller's words become text.
- AI model: what decides the agent's replies.
Each vendor has its own options. These pages say what each option does, its range, and its default, so you can tune an agent with confidence.
| Page | What it covers |
|---|---|
| Voices | Kalpa, Navana, Maya, Gemini and Soniox voices: models, voices, languages and settings |
| Speech recognition | Deepgram, Sarvam, Gnani, AssemblyAI, Gladia, Gemini and OpenAI: each model's languages, turn-taking and settings |
| AI models | The models of each vendor, reply length, creativity and reasoning |
| Language detection | What listens for the caller's language on an agent with several languages |
Where you choose them
- Agent → Voice Settings: the agent's voice and speech recognition. On an agent with several languages, these are the language calls start in.
- Agent → Languages: each language's own voice and speech recognition, with the same options. Open a language and choose Edit.
- Agent → AI Model Settings: the model the agent replies with.
- Flow builder → Settings: Voice & languages for each language's voice and recognition, and Models for the reply and decision models.
A change saves as soon as you make it in Voice Settings and AI Model Settings; in Languages and the flow builder it saves with the language or the flow.
What your platform offers
Kalpa, Navana, Maya, Gemini and Soniox voices, AssemblyAI, Gladia, Gemini and OpenAI recognition, and the models of Google Gemini, Anthropic Claude, OpenRouter, Groq, Together AI and Sarvam are offered only where your platform can run them. The pickers list exactly what yours can use, so whatever you can choose will work on a call. Azure OpenAI models are always available.
Not available on this platform yet
An agent that already uses a vendor your platform can't run keeps it, shown as it is and marked Not available on this platform yet. Its settings are read only, and everything else about the agent still saves as usual. To change it, choose another voice, engine or model.
Languages
A picker offers only what can do the language in question:
- a voice that can't speak the language is shown but can't be chosen, with the reason beside it;
- a recognition engine, or one of its models, that can't recognise the language is shown the same way;
- when you add a language, its voice and recognition start from the agent's own if they can do the language, and otherwise from the standard choice for it.
Each page lists the languages every vendor covers.
Hearing a voice
The play button beside a voice plays a sample. Kalpa, Navana, Maya, Gemini and Soniox voices can't be played from the dashboard yet: their pickers say so. Place a test call to hear one.
When a setting can't be saved
Every field checks its value against the vendor's range as you leave it, and says in plain words what it takes, for example Gladia's pause before a turn ends must be a whole number of milliseconds from 10 to 10,000. Nothing is saved until the value is one the vendor accepts. Leave a field empty to use the vendor's own default, which the field shows.