Q4
01/12/2026, 10:08 AMAI Evy
01/12/2026, 10:08 AMAI Evy
01/12/2026, 10:09 AMAI Evy
01/12/2026, 10:09 AMAI Evy
01/12/2026, 10:09 AMQ4
01/12/2026, 10:12 AMAI Evy
01/12/2026, 10:12 AMAI Evy
01/12/2026, 10:12 AMAI Evy
01/12/2026, 10:12 AMAI Evy
01/12/2026, 10:12 AMSmartb
01/13/2026, 3:03 PMAI Evy
01/13/2026, 3:03 PMAI Evy
01/13/2026, 3:03 PMinterruption_sensitivity parameter controls how easily users can interrupt the agent. [(1) ]()
- Value ranges from [0,1]
- Lower values = harder for users to interrupt (agent waits longer/more words before yielding)
- Higher values = easier for users to interrupt
- Setting to 0 means the agent would never be interrupted
For noisy environments, set it lower (e.g., 0.8) to reduce false interruptions from background speech and noise. [(2) ]()
2. Use Denoising Mode for Noisy Environments
- Remove noise (default) : Removes background noise without impacting transcription accuracy
- **Remove noise + background speech**: More aggressive mode for loud background speech (TV, construction sites) , but may affect transcription accuracy. Incurs $0.005/min surcharge.
3. Handle Interruptions in Your Prompt
Community members have found success with prompt-based handling: [(3) ]()
- When interrupted, have the agent pause and ask "Please go ahead" or "Did you say something?"
- If the user says they didn't say anything, resume from where the agent stopped
- If the user did speak, acknowledge it, repeat back what was said, then ask permission to continue
4. Use no_interruption_allowed for Critical Content
When using Custom LLM via WebSocket, you can set no_interruption_allowed to true in your response events so the agent won't be interrupted during important information. [(6) ]()AI Evy
01/13/2026, 3:03 PMAI Evy
01/13/2026, 3:03 PMSmartb
01/13/2026, 4:15 PMAI Evy
01/13/2026, 4:15 PMAI Evy
01/13/2026, 4:16 PMAI Evy
01/13/2026, 4:16 PMAI Evy
01/13/2026, 4:16 PMAI Evy
01/13/2026, 4:16 PMSmartb
01/13/2026, 4:17 PMSmartb
01/13/2026, 4:19 PMAI Evy
01/13/2026, 4:19 PMAI Evy
01/13/2026, 4:20 PMeleven_turbo_v2, eleven_flash_v2, eleven_turbo_v2_5, eleven_flash_v2_5, eleven_multilingual_v2 (ElevenLabs)
- sonic-2, sonic-3, sonic-3-latest, sonic-turbo (Cartesia)
- tts-1, gpt-4o-mini-tts (OpenAI)
- speech-02-turbo (MiniMax)
Voice Temperature
- Range: [0, 2]
- Lower values = more stable/consistent voice
- Higher values = more variant speech generation
- Default: 1
- Recommendation for clinic: Keep at 1 or slightly lower (e.g., 0.8) for consistent, professional-sounding responses
Voice Speed
- Range: [0.5, 2]
- Lower values = slower speech
- Higher values = faster speech
- Default: 1
- Recommendation for clinic: 1 or slightly slower (e.g., 0.9) to ensure clarity, especially for appointment details
Volume
- Range: [0, 2]
- Lower values = quieter speech
- Higher values = louder speech
- Default: 1
- Recommendation for clinic: 1 is typically appropriate; adjust based on caller feedback
Note: Voice temperature currently only applies to ElevenLabs (11labs) voices.AI Evy
01/13/2026, 4:20 PMAI Evy
01/13/2026, 4:20 PMSmartb
01/18/2026, 2:42 PMAI Evy
01/18/2026, 2:42 PMAI Evy
01/18/2026, 2:43 PMinterruption_sensitivity parameter controls how sensitive the agent is to user interruptions. [(1) ]()
- Range: [0, 1]
- Lower values = takes longer/more words for user to interrupt the agent
- Higher values = easier for user to interrupt the agent
- When set to 0, agent would never be interrupted
- Default: 1
For noisy environments or to reduce false interruptions from background speech, set it lower (e.g., 0.8) . [(2) ]()
2. Set Denoising Mode
Two options available:
- Remove noise (default) : Removes background noise with nearly no waveform distortion, no impact on speech-to-text accuracy
- **Remove noise + background speech**: More aggressive mode for loud background speech (TV, construction sites) . May distort waveform and lower transcription accuracy in some cases. Incurs $0.005/min surcharge.
3. Handle Interruptions via Prompt
Community members have found success with prompt-based handling: [(3) ]() , [(4) ]()
- When interrupted, have the agent pause and say "Please go ahead" or "Did you say something?"
- If user responds "No, I didn't say anything": Resume from where the agent stopped
- If user did speak: Acknowledge with "Thank you for sharing that," repeat back what was said, then ask "May I share what I was saying before we continue?"
- Always ensure the agent resumes the original point instead of jumping ahead [(5) ]()
**4. AdditioAI Evy
01/18/2026, 2:43 PMAI Evy
01/18/2026, 2:43 PMSmartb
01/18/2026, 2:46 PMAI Evy
01/18/2026, 2:46 PMAI Evy
01/18/2026, 2:46 PM## Interruption Handling:
- If a customer makes a noise or interrupts while you are speaking:
- Pause and say:
- "Please go ahead." or "Did you say something?"
- **If they respond with**:
- **"No, I didn't say anything":** Resume from where you stopped in your sentence.
- **"Yes, I was trying to say...":**
- Listen carefully.
- Acknowledge with: "Thank you for sharing that." Then repeat back what they said to confirm with them that you heard them.
- Politely ask: "May I share what I was saying before we continue?"
## Maintaining Context:
- If interrupted, always ensure you resume the original point instead of jumping ahead.
- Avoid skipping key parts of the explanation unless the customer explicitly requests it.
Additional Prompt Tips
- Include backchanneling phrases like "That's great!", "Wonderful!", "Good to know!" to acknowledge user input
- Have the agent ask if it's a good time to talk first, especially in noisy environments
Combine with Settings
Lower the interruption_sensitivity parameter (e.g., 0.8 or lower) to make it harder for background noise to trigger interruptions. [(2) ]() , [(3) ]()
For extremely noisy environments, you may need to experiment with even lower settings, though this will also make it harder for users to intentionally interrupt. [(2) ]()AI Evy
01/18/2026, 2:46 PMAI Evy
01/18/2026, 2:46 PMSmartb
01/18/2026, 3:05 PMAI Evy
01/18/2026, 3:05 PMAI Evy
01/18/2026, 3:05 PMresponsiveness parameter controls how responsive the agent is during conversations. [(1) ]() , [(2) ]() , [(3) ]()
Configuration Details:
- Range: [0, 1] [(1) ]()
- Lower values = less responsive agent (waits more, responds slower)
- Higher values = faster exchanges (responds when it can)
- Default: 1
Recommendation for a Clinic Receptionist:
The documentation doesn't specify a "most recommended" value—it depends on your use case. However, based on the parameter behavior:
- **Default `1`**: Best for natural, quick conversational exchanges where you want the agent to respond promptly
- **Lower values (e.g., `0.7`–`0.9`) **: Better if you want the agent to wait longer before responding, giving users more time to finish speaking
For a clinic receptionist handling appointment details or elderly callers who may speak slower, you might consider lowering responsiveness slightly (e.g., 0.8 or 0.9) to ensure the agent doesn't respond too quickly before the caller finishes. Combine this with lower interruption_sensitivity for best results. [(4) ]()AI Evy
01/18/2026, 3:05 PMAI Evy
01/18/2026, 3:05 PMSmartb
01/19/2026, 3:44 PMAI Evy
01/19/2026, 3:44 PMAI Evy
01/19/2026, 3:44 PM## Interruption Handling:
- If a customer makes a noise or interrupts while you are speaking:
- Pause and say:
- "Please go ahead." or "Did you say something?"
- **If they respond with**:
- **"No, I didn't say anything":** Resume from where you stopped in your sentence.
- **"Yes, I was trying to say...":**
- Listen carefully.
- Acknowledge with: "Thank you for sharing that." Then repeat back what they said to confirm with them that you heard them.
- Politely ask: "May I share what I was saying before we continue?"
## Maintaining Context:
- If interrupted, always ensure you resume the original point instead of jumping ahead.
- Avoid skipping key parts of the explanation unless the customer explicitly requests it.
Additional prompt guidance for long user responses: [(2) ]()
When expecting extended user input (like describing an incident) , include explicit instructions:
## Golden Rule
As you listen to the user's description, you will focus on understanding the context and content of their statement. Only at the end of their statement should you move on to the next step. Under no circumstance should you move to the next step while the user is still describing. If you are not sure whether or not the user has completed their description, you will ask in a polite manner to confirm if you can proceed, before moving to the next step. Never assume that the user has completed their account based on a pause alone; pauses in speech don't necessarily mean the user has completed.AI Evy
01/19/2026, 3:44 PMAI Evy
01/19/2026, 3:44 PMSmartb
01/19/2026, 3:51 PMAI Evy
01/19/2026, 3:51 PMAI Evy
01/19/2026, 3:51 PMinterruption_sensitivity to 0 to prevent the agent from ever being interrupted. [(2) ]()
json
"interruption_sensitivity": 0
3. Prompt for Graceful Recovery
Add this to your prompt to handle any accidental interruptions while keeping flow natural: [(3) ]()AI Evy
01/19/2026, 3:51 PM## Backchanneling:
- When the customer responds to a question or shares information, use one of the following backchanneling phrases to acknowledge their input and keep the conversation engaging:
1. "That's great!"
2. "Wonderful!"
3. "I love hearing that!"
4. "Fantastic!"
5. "Good to know!"
6. "Interesting!"
- Alternate between these phrases throughout the conversation to maintain a natural tone.
## Interruption Handling:
- If a customer makes a noise or interrupts while you are speaking:
- Pause and say:
- "Please go ahead." or "Did you say something?"
- **If they respond with**:
- **"No, I didn't say anything":** Resume from where you stopped in your sentence.
- **"Yes, I was trying to say...":**
- Listen carefully.
- Acknowledge with: "Thank you for sharing that." Then repeat back what they said to confirm.
- Politely ask: "May I share what I was saying before we continue?"
## Maintaining Context:
- If interrupted, always ensure you resume the original point instead of jumping ahead.
- Avoid skipping key parts of the explanation unless the customer explicitly requests it.
Note on Keypad Interruptions
Be aware that Block Interruptions currently only applies to voice input—keypad presses can still interrupt. This has been reported as a feature request. [(4) ]()AI Evy
01/19/2026, 3:52 PMAI Evy
01/19/2026, 3:52 PMSmartb
01/21/2026, 3:16 PMAI Evy
01/21/2026, 3:16 PMAI Evy
01/21/2026, 3:16 PMinterruption_sensitivity and responsiveness, but do not reference "Post-match endpointing" by that name.
Check out these potentially useful sources:
- [Contact Sales]()
- [)
Would you like me to provide a general answer based on my knowledge instead?AI Evy
01/21/2026, 3:16 PMAI Evy
01/21/2026, 3:16 PMSmartb
01/21/2026, 3:28 PMAI Evy
01/21/2026, 3:29 PMAI Evy
01/21/2026, 3:29 PMIf the ID validation returns invalid, say: "The ID number appears to be incorrect. Could you please repeat it?"
The tools need to be defined within the component and will not be visible outside at agent level. [(1) ]()AI Evy
01/21/2026, 3:29 PMAI Evy
01/21/2026, 3:29 PMSmartb
01/21/2026, 3:31 PMAI Evy
01/21/2026, 3:31 PMAI Evy
01/21/2026, 3:32 PMAI Evy
01/21/2026, 3:32 PMAI Evy
01/21/2026, 3:32 PMSmartb
01/21/2026, 3:33 PMAI Evy
01/21/2026, 3:33 PMAI Evy
01/21/2026, 3:33 PM{{variable}} equals 'true'" or uses equation-based transitions for simple comparisons. , [(2) ]()
For your Luhn checksum validation, you would need to:
1. Multiply 9 digits by weights
2. Sum digits of results > 9
3. Sum all values
4. Check if divisible by 10
This multi-step mathematical calculation cannot be performed within a Logic Split Node—it can only evaluate pre-existing values, not compute new ones.
Your realistic options:
1. Custom function (recommended) : External endpoint performs the calculation and returns valid/invalid
2. Prompt-based (unreliable) : Ask the LLM to calculate in a Conversation Node, but LLMs often miscalculate math
If you had a function that returned {{id_valid}} as true or false, then you could use a Logic Split Node to branch based on that result. [(3) ]()AI Evy
01/21/2026, 3:33 PMAI Evy
01/21/2026, 3:33 PMSmartb
01/22/2026, 7:50 AMAI Evy
01/22/2026, 7:50 AMAI Evy
01/22/2026, 7:51 AMmodel_temperature parameter for faster, more deterministic responses. [(1) ]()
json
"model_temperature": 0.7
2. Use Different Models Per Node
Since pricing and performance are calculated based on time spent in each node, you can optimize by using cheaper/faster models for simple routing nodes and premium models only for complex interactions. [(2) ]()
3. Split Complex Nodes
If a node contains too many instructions, the agent may struggle to follow them consistently. Split complex nodes into multiple simpler nodes for more predictable behavior. [(3) ]()
4. Simplify Transition Conditions
Complex transition conditions can slow down decision-making. Break down complex conditions into multiple simpler ones.
5. Review Node Structure
Every node defines a small set of logic, and transition conditions determine which node to transition to. Keep each node focused on a specific task to reduce processing time. [(2) ]()
6. Check Latency in Dashboard
You can view the estimated latency and cost of the agent within the Dashboard. Modify settings there—all changes are auto-saved.AI Evy
01/22/2026, 7:51 AMAI Evy
01/22/2026, 7:51 AMSmartb
01/24/2026, 6:38 PMAI Evy
01/24/2026, 6:38 PMAI Evy
01/24/2026, 6:38 PMbegin_message_delay_ms when creating or updating an agent. [(2) ]()
The begin_after_user_silence_ms parameter is also available in the Conversation Flow API — if set, the AI will begin the conversation after waiting for the user for the duration (in milliseconds) specified. This only applies if the agent is configured to wait for the user to speak first. [(3) ]()
Tip: Enable Pause Before Speaking to avoid cutting users off at the start of calls. [(4) ]()AI Evy
01/24/2026, 6:38 PMAI Evy
01/24/2026, 6:38 PMSmartb
01/24/2026, 7:09 PMAI Evy
01/24/2026, 7:09 PMAI Evy
01/24/2026, 7:09 PMPrompt or Static Sentence.
- **Global Node**: You can set the End Node as a global node so it's accessible from any point in the conversation
Important tip for smooth endings:
If you want a brief pause before the call ends (so it doesn't feel abrupt) , create a Conversation Node before your End Node with your farewell message and enable Skip Response, then connect it to the End Node. [(3) ]()
You can also set the End Node as a global node combined with max call duration and max silence duration settings to ensure the agent eventually hangs up even if the end condition isn't triggered naturally. [(4) ]()AI Evy
01/24/2026, 7:09 PMAI Evy
01/24/2026, 7:09 PMSmartb
01/24/2026, 7:10 PMAI Evy
01/24/2026, 7:10 PMAI Evy
01/24/2026, 7:10 PMPrompt or Static Sentence.
If "Speak During Execution" doesn't work reliably:
A workaround is to move the farewell logic to your prompt instead of relying on the node setting. Instruct the agent to output the farewell sentence first, then immediately trigger the end call—without waiting for user input between the speech and the hangup. [(3) ]()
Best practice for reliability:
You can set the End Node as a Global Node so it's accessible from any point in the conversation. Combined with max call duration and max silence duration settings, this ensures your agent hangs up eventually. [(4) ]() , [(5) ]()AI Evy
01/24/2026, 7:10 PMAI Evy
01/24/2026, 7:10 PMSmartb
01/26/2026, 12:12 PMAI Evy
01/26/2026, 12:12 PMAI Evy
01/26/2026, 12:12 PM{{is_within_work_hours}} that your backend calculates based on current time. [(1) ]()
2. Use conditional logic before transfer
In Conversation Flow, add a Logic Split Node before the Transfer Node to check the {{is_within_work_hours}} variable. Only route to the transfer node if the value is true.
3. Update your prompt with transfer conditions
Include instructions in your prompt such as:
If the user requests to speak with a human and {{is_within_work_hours}} is true, use the transfer_call tool to transfer the conversation to a human agent. If {{is_within_work_hours}} is false, inform the user that our team is currently unavailable and offer to take a message or schedule a callback.
4. Alternative: Handle time logic in a custom function
Create a custom function that checks the current time against your business hours and returns whether transfer is allowed. Use the function result to conditionally trigger the transfer.
Key point: The transfer_call function itself doesn't have built-in time-based restrictions—you need to implement the business hours logic externally (via dynamic variables or a custom function) and use that to control when the transfer is triggered.AI Evy
01/26/2026, 12:12 PMAI Evy
01/26/2026, 12:12 PMSmartb
01/26/2026, 12:16 PMAI Evy
01/26/2026, 12:16 PMAI Evy
01/26/2026, 12:16 PM{{is_within_work_hours}} before routing to the Transfer Node [(1) ]()
2. Transition conditions on the edge leading to your transfer node that evaluate whether it's within business hours
Please share more details or describe what you're seeing, and I can provide specific guidance.AI Evy
01/26/2026, 12:16 PMAI Evy
01/26/2026, 12:16 PMSmartb
01/26/2026, 12:17 PMAI Evy
01/26/2026, 12:17 PMAI Evy
01/26/2026, 12:17 PMretell_llm_dynamic_variables field:
json
{
"user_name": "John Smith",
"product_name": "Premium Plan",
"account_status": "active"
}
*3. For Inbound Calls (Webhook) *
Use the inbound call webhook response with the call_inbound field containing `dynamic_variables`: [(2) ]()
json
{
"call_inbound": {
"dynamic_variables": {
"caller_name": "value"
}
}
}
4. For HubSpot Integration
Pass data using the Dynamic Variables field in JSON format when configuring the call form. Ensure all values are surrounded by quotes. [(3) ]()
5. Test Calls from Dashboard
You can enter dynamic variables when sending a test call directly from the platform. [(4) ]()
Important: For inbound webhooks, make sure to use the correct format with call_inbound wrapping—using retell_llm_dynamic_variables directly in the webhook response will cause an error. [(2) ]()AI Evy
01/26/2026, 12:17 PMAI Evy
01/26/2026, 12:17 PMSmartb
01/26/2026, 12:19 PMAI Evy
01/26/2026, 12:19 PMAI Evy
01/26/2026, 12:19 PM{{is_within_work_hours}} is a **custom dynamic variable you create**—it's not a built-in system variable. You need to calculate it externally and pass it to Retell.
Where to configure it:
For Outbound Calls (API) :
Set it in the retell_llm_dynamic_variables field when calling the [Create Phone Call]() API: [(1) ]()
json
{
"is_within_work_hours": "true"
}
For Inbound Calls (Webhook) :
Return it in your inbound call webhook response: [(2) ]()
json
{
"call_inbound": {
"dynamic_variables": {
"is_within_work_hours": "true"
}
}
}
For Agent-Level Defaults:
Configure default values in the agent settings in the Dashboard. These serve as a fallback when variables aren't included in the call request. [(1) ]()
Important: All values must be strings (e.g., "true" not true) .
Your backend logic should calculate whether it's within work hours before initiating the call or responding to the inbound webhook, then pass the appropriate value.AI Evy
01/26/2026, 12:19 PMAI Evy
01/26/2026, 12:19 PMSmartb
01/26/2026, 12:27 PMAI Evy
01/26/2026, 12:27 PMAI Evy
01/26/2026, 12:27 PM{{is_within_work_hours}} directly within the flow editor—dynamic variables like this are passed into the call externally (via API for outbound calls or webhook for inbound calls) as explained in my previous response.
How to use it within Conversation Flow:
Once the variable is passed into the call, you can use a Logic Split Node to branch based on its value. [(1) ]()
1. Add a Logic Split Node before your Transfer Node [(2) ]()
2. Configure a condition that checks if {{is_within_work_hours}} equals "true" [(1) ]()
3. Connect the matching path to your Transfer Node
4. Connect the else path (when outside work hours) to a Conversation Node that offers alternatives like taking a message
Important note from community feedback:
Logic Split Nodes evaluate existing variable values—they cannot calculate or create values. The variable must already be populated when the call reaches that node. There have been reports of Logic Split Nodes defaulting to the "else" path when variables aren't properly populated before reaching the node. [(3) ]()
Make sure your {{is_within_work_hours}} variable is being passed correctly at call initiation (outbound API or inbound webhook) before the flow reaches the Logic Split Node.AI Evy
01/26/2026, 12:27 PMAI Evy
01/26/2026, 12:27 PMSmartb
01/31/2026, 5:02 PMAI Evy
01/31/2026, 5:02 PMAI Evy
01/31/2026, 5:02 PMAI Evy
01/31/2026, 5:02 PMAI Evy
01/31/2026, 5:02 PMSmartb
01/31/2026, 5:28 PMAI Evy
01/31/2026, 5:28 PMAI Evy
01/31/2026, 5:29 PMgpt-4.1, gpt-4.1-mini, gpt-4.1-nano
- gpt-5, gpt-5.1, gpt-5.2, gpt-5-mini, gpt-5-nano
- claude-4.5-sonnet, claude-4.5-haiku
- gemini-2.5-flash, gemini-2.5-flash-lite, gemini-3.0-flash
For speech-to-speech models: gpt-4o-realtime, gpt-4o-mini-realtime, gpt-realtime, gpt-realtime-mini
Many LLMs are capable of handling multiple languages, enabling global voice AI applications. [(2) ]()
Recommendation: Since the documentation doesn't specify Hebrew performance benchmarks, I'd suggest testing a few models (like gpt-4.1 or claude-4.5-sonnet) with Hebrew conversations to evaluate accuracy and naturalness for your use case. You may also want to reach out to Retell support for specific multilingual guidance.
Check out these potentially useful sources:
- [Choosing the Best LLM for Your AI Voice Agents]()
- [Contact Sales]()
Would you like a general answer based on broader LLM knowledge about Hebrew language support?AI Evy
01/31/2026, 5:29 PMAI Evy
01/31/2026, 5:29 PMSmartb
02/01/2026, 2:53 PMAI Evy
02/01/2026, 2:53 PMAI Evy
02/01/2026, 2:53 PMAI Evy
02/01/2026, 2:54 PMAI Evy
02/01/2026, 2:54 PMSmartb
02/04/2026, 10:45 AMAI Evy
02/04/2026, 10:45 AMAI Evy
02/04/2026, 10:46 AMAI Evy
02/04/2026, 10:46 AMAI Evy
02/04/2026, 10:46 AMSmartb
02/04/2026, 10:48 AMAI Evy
02/04/2026, 10:48 AMAI Evy
02/04/2026, 10:49 AMAI Evy
02/04/2026, 10:49 AMAI Evy
02/04/2026, 10:49 AMSmartb
02/04/2026, 10:50 AMAI Evy
02/04/2026, 10:50 AMAI Evy
02/04/2026, 10:50 AMAI Evy
02/04/2026, 10:50 AMAI Evy
02/04/2026, 10:50 AMSmartb
02/04/2026, 12:52 PMAI Evy
02/04/2026, 12:52 PMAI Evy
02/04/2026, 12:52 PMAI Evy
02/04/2026, 12:52 PMAI Evy
02/04/2026, 12:52 PMSmartb
02/04/2026, 12:57 PMAI Evy
02/04/2026, 12:57 PMAI Evy
02/04/2026, 12:57 PMvoice_speed controls speed of voice with a value ranging from [0.5, 2]
- Lower value means slower speech, while higher value means faster speech rate
- If unset, default value 1 will apply
For Multilingual Agents:
- Set language to multi for multilingual support
- For specific languages like Hebrew (he-IL) , you can set the language parameter directly
Community Feedback on Non-English Languages:
[(2) ]()
If your agent is talking too fast between responses in non-English languages, STT might miss the start of user input. You can adjust the voice response speed in the voice model settings or prompt.
Recommendation:
For multilingual use with non-English languages, consider starting with a slightly lower voice speed (e.g., 0.8-0.9) to help ensure speech recognition catches user input correctly. Test and iterate based on your specific language combinations.
I don't have specific documentation about "Cimo" voice configurations. You may want to reach out to Retell support for voice-specific guidance.
Check out these potentially useful sources:
- [Create Voice Agent]()
- [Contact Sales]()AI Evy
02/04/2026, 12:57 PMAI Evy
02/04/2026, 12:57 PMSmartb
02/04/2026, 12:58 PMAI Evy
02/04/2026, 12:58 PMAI Evy
02/04/2026, 12:59 PMlanguage parameter to multi for multilingual support. Available language options include:
- en-US, en-GB, en-AU, en-IN, en-NZ
- de-DE, es-ES, es-419, fr-FR, fr-CA
- ja-JP, pt-PT, pt-BR, zh-CN, ru-RU, it-IT, ko-KR
- he-IL (Hebrew) , ar-SA (Arabic)
- And many more regional variants
Multilingual Support Update:
[(2) ]()
Multilingual mode now supports seamlessly switching between English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch (using nova 3) .
Key Configuration Settings to Consider:
[(1) ]() , [(3) ]()
- **`voice_speed`**: Value ranging from [0.5, 2]. Lower = slower speech, higher = faster. Default is 1.
- **`responsiveness`**: Value ranging from [0, 1]. Lower = less responsive (waits more) . Reducing by 0.1 adds 0.5 seconds of wait time.
- **`interruption_sensitivity`**: Value ranging from [0, 1]. Lower = more resilient to interruptions and background speech.
- **`boosted_keywords`**: Add commonly used terms, names, or brand words to improve transcription accuracy for domain-specific vocabulary.
Speech Normalization:
[(1) ]()
Enable normalize_for_speech to convert entities like dates, currency, and numbers into spoken form for more consistent speech synthesis.
For specific multilingual voice configurations, I'd recommend testing different combinations and reaching out to Retell support for language-specific guidance.AI Evy
02/04/2026, 12:59 PMAI Evy
02/04/2026, 12:59 PMSmartb
02/04/2026, 2:52 PMAI Evy
02/04/2026, 2:52 PMAI Evy
02/04/2026, 2:52 PMAI Evy
02/04/2026, 2:52 PMreminder_trigger_ms) : If set (in milliseconds) , will trigger a reminder to the agent to speak if the user has been silent for the specified duration after some agent speech. Default value is 10000 ms (10 seconds) . [(4) ]() , [(5) ]()
Retell AI's turn-taking model uses contextual analysis combining sound signals with semantic understanding to determine whether a user is pausing or finished speaking, rather than relying solely on silence duration. [(1) ]()
Could you clarify what specific behavior you're trying to achieve? TAI Evy
02/04/2026, 2:53 PMAI Evy
02/04/2026, 2:53 PM