The dashboard’s Voices, Language, and ASR sections are labelled previews whose changes are not saved. Additional-language first messages and translation are also previews. Use the API workflow below for draft authoring. Voice publication and live conversation execution remain gated.
Before you begin
You need an agent you can update, the intendedbranch_id, a server-side API key with the required access, and a saved voice in the same workspace. To create a complete voice configuration from the start, follow the API quickstart.
Choose and save compatible settings
1
Read the branch and model catalog
Read the selected branch with Get agent. Keep the current document so you can preserve fields you are not changing.Use List available AI models and Get available AI model to inspect available ASR and TTS models. Check language coverage, recognition modes, and supported options instead of copying an identifier from a screenshot.
2
Configure input recognition
Set
conversation_config.asr.model_id to the selected catalog model. Supply its required recognition configuration using a mode that model supports.Choose the caller’s input language independently from the response language. If you add keyterms, use distinct, nonblank terms within the model’s limits. Only supply options that match the selected model’s parameter schema.3
Configure speech generation and response language
Use List my voices to choose an accessible saved
voice_id. Set conversation_config.tts.model_id and conversation_config.tts.voice_id, then check compatibility with the intended languages.Set conversation_config.agent.language for the base response language. Review conversation_config.agent.response_language_policy: base_language is the default; recognized_language requests responses in a language identified by compatible ASR.4
Save and inspect a server draft
Use Create agent draft with an explicit
branch_id and the complete snapshot. Follow versions and drafts and preserve unrelated fields.Read the selected branch’s draft back and confirm ASR mode, model IDs, voice ID, base language, and response policy. Do not use Update agent to publish this voice configuration: the runtime projection is unavailable, even with compatible speech settings.Input-language modes
All fields below are insideconversation_config.asr.recognition.
Selected languages must be distinct. When a
primary_language is supplied, it must appear in languages. Some model contracts require that field; others do not accept it. The presence of a mode in the request schema does not mean every model offers it.
The ASR task currently accepts only transcribe. A response-language policy does not turn speech recognition into a translation task.
Response language and localized copy
conversation_config.agent.language defaults to en. ASR input-language selection does not replace this field.
conversation_config.language_presets can store localized first-message and duration-limit copy, localized soft-timeout settings, and compatible voice overrides. Presets do not switch the ASR model, TTS model, language model, or tool set. Their saved copy does not prove automatic translation or runtime language switching.
Speech fields to review
These paths are relative toconversation_config.
ASR and TTS compatibility is necessary for a future voice publication path, but does not make that path available today. A configuration with
conversation.text_only: true does not require speech settings, but it cannot use recognized_language as its response policy.
Frequently asked questions
Why did my dashboard language or voice selection not save?
Why did my dashboard language or voice selection not save?
The visible speech and language pickers currently hold preview values. Use the API to save a server draft, then inspect that draft. The dashboard’s LLM editor is a separate control; persisting its change still requires an eligible publication.
Can I use a provider's public voice ID?
Can I use a provider's public voice ID?
Use an accessible saved workspace voice returned by List my voices. A public provider identifier is not automatically a saved workspace voice, and model compatibility still needs to be checked.
Why is an automatic recognition mode rejected?
Why is an automatic recognition mode rejected?
Recognition modes and primary-language requirements vary by model. Inspect the selected model’s catalog entry and the returned validation error. Do not assume that every alternative in the public schema is supported by that model.
Does publishing confirm that the voice sounds correct?
Does publishing confirm that the voice sounds correct?
Voice publication is currently unavailable. A future publication would still require a separate supported runtime test to establish audio quality, language detection, and conversation behavior. Keep the assessed version ID with that evidence.