AI providers and models
Which speech, language and realtime AI services a flow can use, how their credentials are supplied, and which choices depend on the plan or on optional deployment features. Browse the feature map or configuration finder.
Availability: Next. Entries that depend on an optional feature, a role or a deployment choice name it in their own Availability line.
Features in this family
- Pipeline modes (unified realtime, modular voice, TTS only, LLM only)
- Configure the AI provider of a flow
- Override the provider on a single step
- OpenAI Realtime
- OpenAI GPT-Live (pilot)
- Azure OpenAI Realtime
- xAI Grok Realtime
- Google Gemini Live
- Pythia Realtime
- Realtime turn detection and transcription settings
- OpenAI Chat language model
- Azure speech recognition
- Inworld speech recognition
- Speech synthesis providers (OpenAI, Azure, Inworld)
- Provider credentials and key references
- Platform-managed default keys and AI defaults
- Custom provider endpoints
- Additional AI providers
- AI-side audio format
Pipeline modes (unified realtime, modular voice, TTS only, LLM only)
A flow either hands the whole conversation to one realtime speech model or chains separate recognition, language model and speech synthesis services, with reduced modes for speech output only or text only.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Flow Designer › Provider editor › Pipeline mode(unified, modular, tts_only, llm_only, PipelineMode) | Pipeline mode and changing configuration |
Read more:
- How it works: 2. Choose the processing mode at the
- Set it up: Pipeline mode and changing configuration
Related features: Configure the AI provider of a flow.
Configure the AI provider of a flow
The provider editor sets the default AI services for a whole flow, including provider type, model, voice and credential reference, and shows only the sections that fit the chosen mode.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Flow Designer › Settings › Provider(providerConfig, realtime.type, llm.type, stt.type, tts.type) | Configure a provider in the editor |
Read more:
- Set it up: Configure a provider in the editor
- Reference: Find a provider setting
- Troubleshoot: Check the configured result
Related features: Pipeline modes (unified realtime, modular voice, TTS only, LLM only), Override the provider on a single step, Provider credentials and key references.
Override the provider on a single step
An individual conversation, speech or language model step can use a different provider or model than the flow default, for example a cheaper model for one classification step.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Flow Designer › Node inspector › Provider override(provider, ProviderRef) | Provider overrides |
Read more:
- Reference: Provider overrides
Related features: Configure the AI provider of a flow, Flow step types.
OpenAI Realtime
A speech-to-speech model from OpenAI handles listening, reasoning and speaking in one connection, with selectable model, voice and speaking speed.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › Realtime Provider › OpenAI Realtime(openai_realtime, gpt-realtime-1.5, alloy) | Provider-specific connection fields |
Read more:
- Reference: Provider-specific connection fields
- Reference: Realtime Advanced Settings
Related features: Realtime turn detection and transcription settings, Provider credentials and key references.
OpenAI GPT-Live (pilot)
A pilot realtime option separates the caller-facing voice from the backend that decides the answer, and runs only on voice servers where the operator has switched the pilot on.
Availability: Available from 0.9.18. Runs only on voice servers where the operator enables the pilot.
For: Application builders, Deployment operators
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › Realtime Provider › OpenAI GPT-Live (Pilot)(openai_live, gpt-live-1, marin, INVALID_LIVE_AUDIO_FORMAT) | GPT-Live (Pilot) |
| Services and containers | Voice service environment(GPT_LIVE_PILOT_ENABLED) | Voice variables |
Read more:
- Set it up: GPT-Live (Pilot)
- Reference: Voice variables
Related features: OpenAI Realtime, OpenAI Chat language model.
Azure OpenAI Realtime
The realtime speech model is reached through an Azure OpenAI resource, with the connection address built from resource, domain and deployment or entered as a full URL.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › Realtime Provider › Azure Realtime(azure_realtime, Resource Name, Domain Suffix, Deployment Name, Full Request URL) | Azure URL configuration |
Read more:
- Reference: Azure URL configuration
- Reference: Provider-specific connection fields
Related features: Realtime turn detection and transcription settings, Custom provider endpoints.
xAI Grok Realtime
The realtime conversation runs on an xAI Grok voice model, with its own key, voice list with preview samples and a language hint for transcription.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › Realtime Provider › Grok Realtime(grok_realtime, grok-voice-latest, grok-transcribe, language_hint) | Provider-specific connection fields |
Read more:
- Reference: Provider-specific connection fields
- Reference: Realtime Advanced Settings
Related features: Realtime turn detection and transcription settings.
Google Gemini Live
The realtime conversation runs on a Google Gemini Live model, configured with a model, a Gemini key and an optional connection address.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › Realtime Provider › Gemini Live(gemini_live, gemini-3.1-flash-live-preview, Realtime URL) | Provider-specific connection fields |
Read more:
- Reference: Provider-specific connection fields
Related features: OpenAI Realtime.
Pythia Realtime
A realtime option for a self-hosted speech-to-speech service that is reached only through a required connection address.
For: Application builders, Integrators
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › Realtime Provider › Pythia Realtime › URL(pythia_realtime) | Provider-specific connection fields |
Read more:
- Reference: Provider-specific connection fields
Related features: OpenAI Realtime.
Realtime turn detection and transcription settings
Optional settings for realtime models tune transcription, noise reduction, voice activity detection, caller interruption, tool choice and reasoning effort.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › Realtime Provider › Advanced Settings(vadType, server_vad, semantic_vad, silence_duration_ms, toolChoice, parallel_tool_calls, whisper-1) | Realtime Advanced Settings |
Read more:
- Reference: Realtime Advanced Settings
Related features: OpenAI Realtime, Azure OpenAI Realtime, xAI Grok Realtime.
OpenAI Chat language model
In modular voice and text-only flows, an OpenAI-compatible chat model generates the replies, with model, base address and sampling parameters set in the provider editor.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › LLM Provider › OpenAI Chat(openai_chat, gpt-4o, Base URL, maxTokens, temperature, presencePenalty, frequencyPenalty) | LLM provider controls |
Read more:
- Reference: LLM provider controls
Related features: Pipeline modes (unified realtime, modular voice, TTS only, LLM only), Override the provider on a single step.
Azure speech recognition
Modular voice flows can transcribe callers with Azure Speech, including recognition language, segmentation and phrase hints.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › STT Provider › Azure STT(azure_stt) | Azure STT controls |
| Services and containers | Voice service environment(AZURE_STT_AUDIO_FLUSH_MS, AZURE_REGION) | Voice variables |
Read more:
- Reference: Azure STT controls
- How it works: Configure modular speech recognition
Related features: Inworld speech recognition.
Inworld speech recognition
Modular voice flows can use Inworld as an alternative speech recognition service.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › STT Provider › Inworld STT(inworld_stt) | Inworld STT controls |
Read more:
- Reference: Inworld STT controls
Related features: Azure speech recognition.
Speech synthesis providers (OpenAI, Azure, Inworld)
Modular voice and speech-output-only flows turn text into speech with OpenAI, Azure or Inworld voices, each with its own voice, speed and format options.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › TTS Provider(openai_tts, azure_tts, inworld_tts) | Configure speech synthesis |
| Services and containers | Voice and API service environment(AZURE_TTS_API_KEY, AZURE_REGION) | Voice variables |
Read more:
- Reference: OpenAI TTS
- Reference: Azure TTS
- Reference: Inworld TTS
Provider credentials and key references
Provider keys are entered as references to team secrets rather than plain values, and fields for custom keys appear only where the team may bring its own credentials.
For: Application builders, Team administrators
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › API Key(openaiDefaultKey, azureDefaultKey, team.openaiDefaultKey) | Credential fields and safe examples |
| Team settings | Settings › Variables › Set Secret(Set Secret) | Platform-managed defaults |
Read more:
- Set it up: Credential fields and safe examples
- Troubleshoot: 3. Check credentials and audio requirements
Related features: Platform-managed default keys and AI defaults, Custom provider endpoints.
Platform-managed default keys and AI defaults
The platform administrator registers shared OpenAI and Azure keys that every team receives as managed defaults, and sets the default models used by platform AI features.
For: Deployment operators, Team administrators
| Level | Where to change it | Owner section |
|---|---|---|
| Platform settings | Admin Settings › Platform Settings › AI defaults(openaiDefaultKey, azureDefaultKey, llmDefaults, AZURE_OPENAI_BASE_URL) | Configure AI defaults and check their consumers |
| Team settings | Settings › Variables › platform-managed rows(openaiDefaultKey, azureDefaultKey) | Platform-managed defaults |
Read more:
- Set it up: Configure AI defaults and check their consumers
- Reference: Platform-managed defaults
Related features: Provider credentials and key references.
Custom provider endpoints
A provider can be pointed at another compatible host, such as a regional Azure resource or a proxy, by overriding its base address or speech paths.
For: Application builders, Integrators
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › Advanced Settings › Custom URL / Base URL(customUrl, baseUrl, endpointUrl) | Verify a custom provider connection |
Read more:
- Set it up: Verify a custom provider connection
- Reference: Azure Speech paths and source-only overrides
Related features: Azure OpenAI Realtime, Provider credentials and key references.
Additional AI providers
An optional deployment feature widens the choice of modes, realtime providers and transcription models beyond the default set offered to customers.
Availability: Requires the Additional AI providers feature. Applies to development deployments only; provider support and the team plan stay separate conditions.
For: Deployment operators, Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Portal and feature profiles | Portal › Feature profile › Additional AI providers(additionalProviders) | Integration capabilities |
| Services and containers | Web service environment(FEATURE_ADDITIONAL_PROVIDERS) | Web UI feature conditions |
Read more:
- Reference: Web UI feature conditions
- Reference: See which features are enabled
Related features: Pipeline modes (unified realtime, modular voice, TTS only, LLM only), Configure the AI provider of a flow.
AI-side audio format
The audio format exchanged with the AI service is chosen per direction and filtered to what the selected provider supports, separately from the telephone codec.
For: Application builders
| Level | Where to change it | Owner section |
|---|---|---|
| Flow and node settings | Provider editor › Audio › AI Input / AI Output / Rate(ai.input, ai.output, pcm16, ulaw, alaw, g722) | AI audio format and preprocessing |
Read more:
- Reference: AI audio format and preprocessing
Related features: Voice codecs on SIP calls.