Jarvis Now Speaks Uzbek More Naturally

Jarvis Now Speaks Uzbek More Naturally

2 min read 5 views
Case Studies
Children practice speaking Uzbek with Jarvis and hear responses in the Saodat voice. The new solution shortened the silence before the acknowledgment cue.

Jarvis can now understand children speaking Uzbek and respond in the Saodat voice. The client uses this solution to help children practice their speech. The project’s goal was to replace OpenAI GPT-Live with NeuronAI while preserving the existing live conversation experience.

Once the child finishes speaking, Jarvis first plays a short sound to show that it heard them, then begins its response. The wait before this sound was reduced from 2,25 seconds to 1,28 seconds. The actual response begins approximately 1,5–1,7 seconds after the child stops speaking.

There Used to Be 2,25 Seconds of Silence After the Child Finished Speaking

In speech practice, a good answer alone is not enough. If Jarvis remains silent for too long after the child finishes speaking, it can feel as though the conversation has been interrupted.

That is why the team did not focus on the response content alone. They also improved the timing of both the short acknowledgment sound and the response that follows it.

Jarvis Hears Uzbek and Responds in the Saodat Voice

Jarvis was migrated from OpenAI GPT-Live to NeuronAI. The new solution recognizes Uzbek speech, prepares a response, and delivers it in the Saodat voice. If the primary response path fails, a fallback path takes over.

The previous service was discontinued, but the option to return to it when necessary was retained. Approximately 123 engineering hours were allocated to the project.

The team also tested publicly available, ready-made solutions that convert Uzbek speech directly into a spoken response. In the products they found, the process was divided into three parts: listening to the speech, preparing a response, and reading it aloud.

A test monitor shows the gap between the child's speech, acknowledgement, and reply
Before optimization, the acknowledgement sound was preceded by a 2.25-second pause.

For the Child, the Conversation Happens in Four Simple Steps

The complex process remains invisible to the user. For the child, it is still just a simple conversation.

  1. The child speaks Uzbek Jarvis listens to the child and recognizes the Uzbek words they say.
  2. Jarvis signals that it heard the child Once the child finishes speaking, a short acknowledgment sound plays in the Saodat voice.
  3. The response is prepared If the primary path fails, Jarvis prepares the response through a fallback path.
  4. Saodat delivers the response The completed response is delivered to the child in Uzbek, using the Saodat voice.

The Waiting Time Was Clearly Reduced

2,25 secondsPrevious acknowledgment cue
1,28 secondsNew acknowledgment cue
1,5–1,7 secondsActual response start time
A microphone, chime, voice module, and speaker form the four conversation steps
Four steps: the child speaks, Jarvis acknowledges, processes the speech, and replies.

Three Lessons for Owners of Similar Products

This project offers three practical takeaways for product owners working on voice-based conversations.

  • Measure the silence too. Users notice not only the quality of a response, but also how long they had to wait for it.
  • Add an acknowledgment cue. A short sound immediately lets the child know that their speech has been received.
  • Keep a way back. When replacing a service, retaining the option to return to the previous one reduces technical risk.

Make Voice Conversations Feel More Natural

If your product needs voice conversations in Uzbek, faster responses, or a clear acknowledgment cue, contact Celion.

Contact us

Share this article

You might also like