Voice & Audio Settings

Voice and audio settings determine how your agent sounds, how it understands speech, and which language model powers its reasoning. Getting these right is essential for creating a natural, responsive conversational experience.

Text-to-Speech (TTS)

The TTS provider converts your agent's text responses into spoken audio. Cloudax Connect supports multiple providers, each offering distinct voice characteristics and performance profiles.

Voice Library

The voice library includes hundreds of pre-made voices across different accents, ages, and styles. Browse and preview voices from the Personalities page, then select one in your agent's settings. You can also use a custom voice ID if you have created a custom voice.

Voice Parameters

Depending on the provider, you can adjust the following parameters to fine-tune how your agent sounds:

Stability: controls how consistent the voice sounds across utterances. Higher values produce more uniform delivery; lower values introduce natural variation.
Similarity boost: how closely the output matches the original voice sample. Higher values sound more like the reference voice but may reduce expressiveness.
Style: adjusts the emotional expressiveness of the voice. A higher style value makes the agent sound more animated and engaging.
Speed: controls the speaking rate. Values above 1.0 speed up delivery; values below 1.0 slow it down.

TTS Model

The voice runs on one of several ElevenLabs model tiers, which trade latency against audio quality and language coverage. These are the options in the model dropdown in your agent's voice settings:

Flash 3: the most expressive real-time model, with context-aware delivery and support for more than 70 languages. Choose it when natural emotion and conversational nuance matter most.
Flash 2.5: fast and efficient voice generation with the lowest latency, and supports 32 languages. The best default for snappy, high-volume agents.
Flash 2: the previous-generation Flash model. It is English-only, but tends to produce fewer voice artefacts, so it can be a strong choice for English agents where audio cleanliness matters more than language coverage.
Turbo: higher quality and only slightly slower than Flash, with support for 32 languages. It delivers better audio fidelity and reduces artefacts such as glitches, clipped words, robotic warbling, and mispronunciations.

Hearing audio artefacts?

If your agent's voice occasionally glitches, warbles, clips words, or mispronounces names and numbers, try a different model in the agent's voice settings. For multilingual agents, move from Flash 2.5 to Turbo for better fidelity. For English-only agents, Flash 2 often has fewer of these artefacts.

Choosing a voice

Preview voices on the Personalities page before selecting one. Listen for clarity, naturalness, and how well the voice fits your agent's personality and use case.

Speech-to-Text (STT)

The STT provider transcribes the caller's speech into text so your agent can understand and respond. Accurate transcription is critical for conversation quality.

STT Models

Choose the model that best matches the language(s) your callers will speak:

Conversational (recommended): supports 10 languages (English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch). Pick a single language for best accuracy, multiple languages if your callers switch between them in the same call, or none to auto-detect.
Other languages: supports 50+ languages. Use this when your caller's language isn't one of the ten covered by Conversational.

STT Settings

Language: set the expected language of the caller. This improves transcription accuracy, especially for non-English calls.
Keyterms: provide custom vocabulary (product names, brand terms, industry jargon) to help the STT engine recognise words it might otherwise miss or misspell.
Endpointing delay: the amount of silence (in milliseconds) before the STT engine considers the caller's turn complete. A shorter delay makes the agent more responsive; a longer delay reduces the chance of the agent interrupting the caller mid-thought.

Reduce interruptions

If your agent frequently interrupts callers, try increasing the endpointing delay to 300–500ms. This gives callers more breathing room between sentences.

Tell the LLM it's multilingual

Selecting a multilingual STT model only handles transcription. To make the agent actually speak and understand multiple languages, you must also instruct the LLM in your system prompt that it is multilingual. Without this, the model will tend to reply in a single language regardless of what the caller speaks. Open the agent's system prompt and add a line such as:

You are a multilingual customer service representative. Detect the language the caller is speaking and respond fluently in that same language. If the caller switches languages mid-conversation, switch with them.

Language Models

The language model (LLM) is the brain of your agent: it interprets the caller's intent, follows your system prompt, and generates responses. Available models, listed roughly from newest to oldest:

5.3 Chat: a fast conversational model from the 5-series. Strong general capability with low latency, well suited to natural back-and-forth dialogue.
4.1: the full-size 4.1 model. Use when the agent needs maximum instruction-following accuracy and you can afford a little more latency and cost than the Mini.
4.1 Mini (default) - excellent balance of speed and intelligence. The current recommended starting point for new agents.
4.1 Nano: the fastest model. Ideal for simple, high-volume interactions that don't require complex reasoning.
4o: multimodal model with strong audio-aware capabilities for agents that need particularly natural turn-taking.
4o Mini: multimodal capabilities at a compact size. Suitable for agents that process both audio and text inputs on a tighter latency budget.

Bigger, more capable models (the full-size 4.1 and 5-series models sit at the top of the family, 4.1 Mini in the middle, Nano variants at the bottom) trade extra cost and latency for better reasoning and fewer hallucinations. Most agents do well on 4.1 Mini; reach for a larger model when the agent needs to handle complex tool chains, long policies, or particularly tricky callers.

Temperature

The temperature setting controls how creative or deterministic your agent's responses are. A value of 0 produces highly consistent, predictable responses. A value of 1 introduces more variation and creativity. For most telephony use cases, we recommend a temperature of 0.2 to keep responses focused and on-script.