Model, voice, speech, and transcription settings

These settings control how an AloAi Voice Agent thinks, sounds, and listens. The AI model, voice, transcriber, and language selectors sit at the top of the Configure tab, above the instructions box.

Speech Settings and Realtime Transcription Settings are panels on the right side of the same tab. Each setting applies to one agent, so you can tune a support agent, a booking agent, and an outbound campaign agent differently.

To get there, open the Agents menu and click the agent’s name. The model, voice, transcriber, and language selectors are at the top of the Configure tab.

The model selector sets the AI model the agent uses on calls. Choose a faster model when response speed matters, or a more capable model when the agent needs to handle in-depth conversations.

Click the model selector to open Select Model. Models are grouped into Versatile and highly intelligent, Fast and cost-efficient, and Speech to speech.

Each model shows its tier (Basic, Premium, or Ultra Premium), its provider, and its type (Traditional, Speech-2-Speech, or Full Duplex).

Recommended options appear at the top of each group with a Recommended tag. You can search by model name or provider, or filter the list by Provider and Type.

Available models include:

The recommended options, tiers, and groups change as models are added or retired, so treat the live selector as the authoritative list.

The voice and the AI model are separate settings. Changing the model keeps the voice you selected.

Temperature controls how much the agent's wording varies from one conversation to the next. A lower value keeps the agent close to the same phrasing every time, which suits a scripted intake flow. A higher value lets the wording vary more.

To set the temperature, click the settings button next to the model selector. Adjust LLM Temperature, then click Save. A lower value yields better function call results.

Some models accept a range of temperature values, and others accept only one fixed value. The temperature control appears only for models that accept a range.

For a model that requires a fixed value, the control is hidden and the required value is applied for you. For example, Claude Sonnet 5 requires a temperature of 1.0, so an agent on Claude Sonnet 5 runs at 1.0 and there is nothing to set.

When you switch models:

Note the current temperature before you switch models, so you can set it again if you switch back.

The same window has Structured Output. Turn on Enable Structured Output to make the agent always generate responses that follow the JSON schema you supply.

The voice selector sets how the agent sounds. Click it to open Select Voice, which lists the voices compatible with the agent's model.

Each voice shows its provider, gender, accent, and languages. You can search by name, provider, or accent, or filter the list by Provider, Gender, Accent, and Language. Click the play button next to a voice to preview it before selecting it.

Selecting a voice for one agent does not change the voice on any other agent.

To adjust the voice, click the settings button next to the voice selector. Choose the Voice Model, for example Eleven Flash v2.5, and set Voice Speed, Voice Temperature, and Voice Volume. Click Save to apply the changes.

Voices come from Deepgram, ElevenLabs, Google, OpenAI, and xAI. ElevenLabs voices in English include:

The list also includes Spanish voices with a Latin American accent: Andrea - Spanish, Cheerful and Calm, Monica - Spanish, Friendly & Polite, and Runa - Spanish (Venezuelan). Voices are added over time, so treat the live selector as the authoritative list.

The language selector sets the language the agent speaks, for example English (US). You can search the list by name. Supported languages include:

The transcriber selector sets the speech-to-text model the agent uses to understand callers, for example Deepgram Flux (Multilingual). To fine-tune how it transcribes, see Realtime transcription settings below.

Speech Settings control how the agent sounds and how it manages timing during a conversation. Changes save with the agent and apply to future calls, not calls already in progress.

Some options depend on the voice assigned to the agent. If a setting is disabled or shows a message, the current voice does not support it.

Some functions take a few seconds to finish, such as checking calendar availability, booking an appointment, or searching the knowledge base. While a function runs, the agent keeps the conversation going instead of going silent.

For example, it can say "Let me check the calendar, one moment" and give the answer as soon as the result is ready.

If a lookup takes longer, the agent adds short filler phrases so the caller knows the call is still connected. Filler phrases stop as soon as the result is ready, and quick lookups do not get them. The caller can still interrupt at any time. No setup is required.

The pronunciation dictionary tells the agent how to say specific words, such as brand names, acronyms, and technical terms. For example, you can make sure "Aloware" is spoken correctly instead of being read literally.

The dictionary works only with an ElevenLabs voice and the Eleven Flash v2 voice model, which you select under Voice Model in the voice settings. The Pronunciation section appears only with this combination. It does not appear for other ElevenLabs models (Eleven Flash v2.5, Eleven v3, and Eleven Multilingual v2), for other voice providers, or for speech-to-speech models such as OpenAI Realtime. If the agent is switched to an unsupported model, saved entries are ignored.

Each entry has three fields:

For example, "Aloware" is written as æloʊwɛr in IPA or AE L OW W EH R in CMU. To find the phonemes for a word, use an online converter, such as tophonetics.com for IPA or the CMU Pronouncing Dictionary for CMU, and paste the result into Phoneme.

The editor warns you as you type if the phoneme does not match the selected alphabet. If you see a warning, switch Alphabet to match.

Each agent can have up to 100 entries. If the same word is added more than once, only the last entry applies, and the editor shows an alert.

Realtime Transcription Settings control how the agent converts the caller's speech to text during a call.

Denoising Mode filters out unwanted background noise. Remove noise is selected by default.

Transcription Mode balances speed and accuracy. Optimize for speed is selected by default.

Boosted Keywords expands the model's vocabulary with terms such as your business or product names. Enter them separated by commas, for example Blowfish, Aloware. Keep the list to the key terms.