Feature Requests

I have many conversations where the voice changes completely as it's speaking, change of tone, accent, gender. Making it un-usable for our purposes. Here's a sample chat ID: c66e3ff2-c2dd-42ac-9d52-0c7b2f10bdbc fa0a850c-0c98-49db-8ea6-5bbf955b5c06 e95e9a8c-9189-45ad-a9dd-1ebb2a73e691

Right now when using the Twilio integration for outgoing calls, I can only pass a config_id and that's it. That's very limiting since this does not allow me to start the call with a personalized greeting. It means that I either need to create one config per call (which seems to be not the intention of configs) or send session_settings when I get the CHAT_START_MESSAGE webhook notification. The latter approach has a bit of a delay, so I still cannot use a personalized greeting with a generic config. I think it would be great if we could pass additional data through the Twilio webhook, like dynamic variables. That would help a lot.

The Typescript SDK allows: modelResource: "grok-4-fast-non-reasoning-latest", in hume.empathicVoice.configs.createConfig -> languageModel. But you are not allowed to specify modelProvider: "X_AI", which is necessary for using it.

This feature request is for session_settings or Config to have a parameter that would allow the user to configure EVI's end-of-turn detection to make it error more or less on the side of thinking the user is done speaking and starting to take its turn. Some users have reported that they would like EVI to be a little less eager to start speaking and should allow the user to hesitate a little bit more before jumping in.

User @shakir on discord ( https://discord.com/channels/991790422433210459/1417390598750470256/1418341250733576273 ) reported that EVI was hallucinating headlines even with Web Search enabled. Investigation revealed that the hallucinations originated from the short responses model and were fixed when quick responses were disabled. One possible thing that would help users if more explicit control over the short response prompt. That way the short response model could be instructed to "wait" and not attempt to respond to certain types of queries, to prevent hallucinations but still create a conversational experience.

EVI is failing to STT emails correctly. Any email with two words split with dot before @ have the dot removed, so first.second@gmail.com becomes second@gmail.com . Worse still, the eLLM fails to capture even that correctly, more often than not repeating back something somewhat different, like second@email.com , presumably because it recognises email more than gmail as a word? Is this a known fault? Is there any effort to fix it?

Right now, if you register a tool in your EVI config, EVI will send a tool call message over the websocket and expects the client to send back a tool response message. put a tool call message in the transcript There is no way to intercept tool calls if you're connected via the /v0/twilio webhook. Even if you can intercept tool calls, you are forced to implement the logic for calculating the tool response in the client, which might not be the most convenient place -- you might want your server to be able to handle the tool call, or to be sending it some place like Zapier. Specific phone call related use cases that this would unlock: - Transferring the caller (on a phone call) - Scheduling an appointment during the call - Sending a text message during the call

I am retrieving chats with the API but some are just tests or were unsuccessful and I don't want to include them in my reporting.

We are using Hume AI in an industry with strict data privacy protection requirements. I would like an API endpoint that allows us to delete chat sessions after use or an auto-deletion setting in Hume AI that removes chat history after 1 or 2 days.

EVI easily picks up background noise and sounds (chatter, music, etc) which causes a lot of disruption to the conversation. Feature request for background noise filtering and speaker isolation. I tried doing so using the prompt but it didn't help. It doesn't pause the conversation, but the EVI is "listening" to the background and takes that as part of the conversation now