Skip to main content

Conversation types

There are two conversation types that determine how the voice agent processes speech and responds to callers:
  • Standard Conversation - a three-stage process (STT → LLM → TTS)
  • Realtime Conversation - a direct audio-to-audio approach
Default: When you create a new voice agent, Standard Conversation is used as the conversation type automatically. The mode can be changed at any time under Advanced in the agent’s settings.

Standard Conversation (STT → LLM → TTS)

With a Standard Conversation, every message from the caller passes through three separate processing steps: 🎤 STT → 🧠 LLM → 🔊 TTS
  1. STT (Speech-to-Text) - The caller’s spoken language is converted into text. The speech recognition used delivers a highly precise transcription - even for technical terms, proper names, and dialects.
  2. LLM (Large Language Model) - The recognized text is sent to an AI language model, which generates an appropriate response. The available models support a significantly larger context window than realtime models. This allows the agent to take longer conversation histories, extensive knowledge bases, and complex instructions into account at the same time.
  3. TTS (Text-to-Speech) - The generated text response is converted into spoken language and read to the caller.
In Standard mode, you can fine-tune both speech recognition and speech output: for the STT component, special terms and background context are available to improve recognition accuracy. For the TTS component, you can choose the voice and configure its properties such as speaking rate, stability, and similarity. This mode is particularly well suited for complex use cases with extensive business logic, tool calls, knowledge bases, and multi-step conversation flows.

Choosing a language model

In the Voice Wizard, open Advanced and then Conversation and Audio. Use the Model field to choose the language model for a Standard Conversation: The choice controls the agent’s responses. Fixed millisecond values cannot be stated: prompt length, tool calls, network conditions, and the speech recognition and output pipeline all affect perceived response time. Configure speech recognition and speech output separately.
GPT-5.6 Terra and GPT-5.6 Luna are available only for Standard Conversation. Realtime Conversation uses its own native realtime models.

Separate from Bring Your Own LLM

The Model field selects platform-managed Standard models. It is separate from Bring Your Own LLM; choosing one of these models does not change its connection or behavior.

Realtime Conversation

A Realtime Conversation uses an end-to-end speech model that processes audio directly - without the detour through separate STT and TTS steps: 🎤 Audio in → 🧠 Realtime speech model → 🔊 Audio out The speech model receives the caller’s audio data directly and produces the spoken response in a single step. There is no separate intermediate text stage. The response is nearly instantaneous, which creates a particularly natural and fluid conversation experience. This mode is particularly well suited for simple, straightforward use cases - e.g. FAQ hotlines, general inquiries, or short informational calls where fast response time matters more than complex conversation logic. Speech recognition optimization (special terms, background context) is not available in this mode.

Speech recognition optimization

Speech recognition optimization helps the voice agent better understand the caller’s spoken language. It is available only in Standard Conversation mode and consists of two areas:
  • Special terms - Individual words or short expressions that are explicitly made known to the speech recognition.
  • Background context - Descriptive text that conveys the topic and context of the conversation to the speech recognition.
Character limit: Special terms and background context share a combined budget of 9,000 characters. If both fields together exceed this limit, an error message is displayed.

Special terms

Special terms (also: key terms) are individual words or short expressions that come up frequently in conversations but may be unusual, rare, or unknown to the speech recognition. Registering these terms significantly improves recognition accuracy for exactly these words.

When should you use special terms?

  • Technical terms: Industry-specific words that rarely occur in everyday language
  • Product and brand names: Proper names that are not part of the standard vocabulary
  • Street and place names: Local names the speech recognition may not know
  • Personal names: Names of employees or contacts that are mentioned frequently
  • Abbreviations and acronyms: Company-internal designations or industry-standard abbreviations
  • Invented or compound words: Terms created specifically for your company or your products

Input format

Enter the terms as a comma-separated list:
  • Insurance industry: Schadensmeldung, Deckungssumme, Haftpflicht, Kaskoversicherung, Selbstbeteiligung, Policennummer
  • Medical field: Ibuprofen, Metformin, Pantoprazol, Blutdruckmessgerät, HbA1c-Wert
  • IT service provider: Kubernetes, Terraform, GitLab, On-Premise, Firewall-Regel, VPN-Tunnel
  • Real estate: Grundbuchauszug, Nebenkostenabrechnung, Mietkaution, Energieausweis, Wohnflächenberechnung
  • Company names and products: TarifPlus, SmartConnect, FlexiCare, MediCheck, ProfiPlan

Best practices

  • Only add terms that the speech recognition actually misrecognizes or where you expect misrecognition.
  • Use the correct spelling - the speech recognition uses it as the reference for the text output.
  • Avoid full sentences - special terms should be individual words or short expressions (max. 2–3 words).
  • After adding terms, test whether recognition has improved.

Background context

The background context gives the speech recognition additional information about what the conversations are about. Unlike special terms, which target individual words, the background context provides thematic orientation - similar to telling a listener in advance what is about to be discussed.

When should you use background context?

  • When conversations concern a specific subject area (e.g. insurance, medicine, technology)
  • When typical conversation flows are predictable
  • When there are frequently recurring topics that can help the speech recognition interpret ambiguous words correctly

What should the background context contain?

  • Description of the area of use: What is the voice agent’s task?
  • Typical conversation topics: What do callers usually talk about?
  • Industry-specific information: Which subject areas or products are relevant?

Examples

Best practices

  • Write the context in natural language - as if you were explaining the typical conversation content to someone.
  • Keep the text concise and relevant - only information that actually helps the speech recognition.
  • Mention typical formats (e.g. ticket numbers, case numbers) that callers frequently say.
  • Update the context whenever the area of use of your voice agent changes.
Important: The background context is used exclusively by the speech recognition (STT). It does not influence the behavior or the responses of the AI language model (LLM). To control the agent’s behavior, use the AI instructions / prompts.

Background audio

With background audio, you can give the voice agent a subtle ambient sound that plays in the background throughout the entire conversation. This gives the caller the impression of speaking with a real person in a real environment.

Available sounds

The volume of the background audio can be adjusted continuously with a slider.

Tool execution sound

The tool execution sound is played while the voice agent executes a tool call - e.g. an API query, sending an SMS, or booking an appointment. During this time, the agent is waiting for the external system’s response, and without this sound the caller would hear only silence. The sound signals to the caller that the agent is actively working and bridges the waiting time.

Available sounds

  • Off - No sound during tool execution (default)
  • Keyboard Typing - Keyboard sound, gives the impression that the agent is entering data or looking something up
  • Keyboard Typing 2 - Alternative variant of the keyboard sound

Example

Caller: “Can you tell me whether my contract is still active?”Agent: “One moment please, I’m checking your details…”While the agent queries the API, the caller hears a keyboard sound in the background.Agent: “Your contract with the number V-2024-00815 is active and runs until December 31, 2025.”
Note: The tool execution sound is played only during active tool calls and stops automatically as soon as the result is available. It does not replace the background audio - it complements it.