Customize your Voice
On top of that shared AI Agent configuration, you can adjust settings specific to the Phone channel.
General Tab
Voice Selection
The General tab lets you choose the voice your AI Agent uses when speaking to callers.
Voices are organised by provider (Google or ElevenLabs) using tabs at the top.
| Field | Description |
|---|---|
| Provider tab | Switch between Google Voices and ElevenLabs Voices. |
| Search | Filter voices by name or voice ID. |
| Gender filter | Narrow the list by gender. |
| Voice list | Click any voice to select it. The selected voice is used by the AI Agent for all outbound speech. |
Preview Panel
The Preview panel (visible alongside the General tab) lets you audition the currently selected voice before saving.
Playback uses a live API call to the TTS service. If the call fails due to a network error, invalid configuration, or service unavailability, an error notification is shown and no audio plays.
Advanced Tab
How your AI Agent speaks
Controls the characteristics of the AI Agent's synthesised voice output.
Speed
- Range: 0.25× – 2× (Google) · 0.7× – 1.2× (ElevenLabs)
- Default: 1×
- Controls how fast the AI Agent speaks. Values below 1× slow it down; values above 1× speed it up.
Volume
- Range: 0.5× – 1.5×
- Default: 1×
- Scales the amplitude of the AI Agent's audio output. This is applied on top of the base voice volume, not instead of it.
Higher volume values may introduce audible crackling. Adjust carefully.
Stability (ElevenLabs only)
- Range: 0 – 1 (Expressive → Consistent)
- Default: 0.5
- Controls how consistent the voice sounds across utterances. Higher stability means the voice stays close to its baseline tone; lower stability introduces more expressive variation. Very low values can sound erratic; very high values can sound flat.
Similarity boost (ElevenLabs only)
- Range: 0 – 1 (Loose → Strict)
- Default: 0.75
- Controls how closely the generated speech resembles the original voice model. Higher values reproduce the model's tonal characteristics more faithfully but may amplify artefacts such as background noise or compression. Lower values allow the model more freedom.
Custom vocabulary
A list of phrase → pronunciation pairs that control how the AI Agent pronounces specific words — useful for brand names, acronyms, or technical terms the text-to-speech model renders incorrectly by default.
Each entry has a Phrase (the word as it appears in the script) and a Pronunciation — an alternate spelling the model is more likely to pronounce correctly.
Custom pronunciation is supported for Google voices only. ElevenLabs does not currently support custom vocabulary.
Tips for writing pronunciations
Text-to-speech is AI-driven, so even small spelling changes can meaningfully affect output. Start with the simplest change and iterate:
- Respell or split the word (e.g. "FAQ" → "F A Q", "GIF" → "jiff")
- Try uppercase or spaced-out letters (e.g. "BMW" → "Bee Em Double-you")
- If the word appears at the very beginning or end of a sentence, test the output carefully — the model can behave unexpectedly at sentence boundaries. If problems persist, consider restructuring the script so the term does not open or close a sentence.
How your AI Agent listens
Controls how the AI Agent performs Automatic Speech Recognition (ASR) — i.e., how it converts the caller's voice into text.
Priority phrases
A list of customer- or domain-specific words and phrases the speech recognition model should prioritise — for example, company names, product names, or proper nouns callers are likely to say (e.g. "Microsoft", "Nike").
- Providing these phrases boosts their recognition probability without blocking other words from being recognised.
- If a term is still misrecognised after adding it here, use Post-process replacements as a fallback.
Post-process replacements
A list of find-and-replace rules applied to the caller's transcribed speech after speech recognition, before the text reaches the AI Agent's reasoning layer.
This is the recommended fallback when priority phrases alone are not enough. If the speech recognition model consistently misrecognises a term, add the incorrect output as the From value and the correct form as To.
Use this to:
- Fix persistent misrecognitions that priority phrases did not resolve (e.g. "flower" → "flour", "Sam sung" → "Samsung")
- Correct systematic errors (e.g. a model that consistently writes "to" when the caller says "two")
- Normalise formatting (e.g. replacing "dot" with ".")
- Redact or mask specific terms
Each entry has:
| Column | Description |
|---|---|
| From | The word or phrase to find in the transcription. Matching is case-insensitive and respects word boundaries — partial matches within longer words are not replaced. |
| To | The replacement word or phrase. Can be empty to effectively delete the matched text. |
Post-process replacements run on every transcription from the selected speech recognition provider.
Additional Options
Miscellaneous audio and session settings.
Background noise
Adds a looping ambient audio track mixed into the AI Agent's voice output. This makes the agent sound more natural and human-like — particularly useful when callers expect to hear some call-centre ambience.
| Setting | Description |
|---|---|
| Type | Choose from: Call center 1–5 (varying intensities of call-centre background noise) or Office ambience (quieter, open-office sounds). |
| Volume | Range 0–1. Controls the amplitude of the background audio relative to the agent's voice. |
Background noise is added to the AI Agent's outgoing audio only. It does not affect what the agent hears from the caller.
Beep
Toggleable. When enabled, plays a short confirmation tone after the caller finishes speaking — similar to the "beep" at the end of a voicemail prompt. This signals to the caller that their input was received and the agent is processing.
DTMF (Dual-tone multi-frequency)
Toggleable. When enabled, the system recognises keypad presses (touchtone inputs) during a call. This allows callers to navigate menus or enter numeric data by pressing phone keys instead of speaking.
DTMF is only active during Data Gatherer steps — i.e. when the AI Agent is collecting information from the caller. Keypad input outside of those steps is ignored.
DTMF input is processed in parallel with speech recognition. Both can be active simultaneously. Enabling this does not disable voice input.
Record Session
[Additional permission required]
Controls whether calls are recorded and stored. Users without this permission can see the current setting but cannot change it.
Recordings are retained for 30 days and permanently deleted after that period.
Legal notice: Enabling call recording typically requires informing callers that the conversation is being recorded. Ensure the AI Agent's script includes an appropriate disclosure before enabling this feature. A confirmation prompt is shown when enabling to acknowledge this requirement.