Uzbek speech to text works in production on clean speech and standard phrases: a base model handles those out of the box. It breaks on phone audio, on switching between Uzbek and Russian inside one sentence, on dialects, names and amounts. The cause is data: open corpora contain read speech, not calls; the Uzbek Speech Corpus holds 105 hours from 958 speakers, and its authors report a 17.4% word error rate on the test set. The fix is not a bigger model but data: labelled real calls, a glossary of terms and handling of Uzbek ↔ Russian code-switching. Latency is a separate task: a voice agent is a pipeline of recognition, understanding, response and synthesis, and streaming recognition instead of waiting for the end of a phrase removes the most noticeable pause. Zukko.AI, an AI agent platform for business, runs its own speech recognition models for Uzbek, Russian and English, and its voice AI agent and call analytics are built on them.
Does Uzbek speech to text work in production
A voice agent for a contact centre is not “a model that dials people”. It is a pipeline: speech recognition, request understanding, response and synthesis — all of which must fit within a latency where a human feels no pause. In Uzbek the bottleneck is recognition itself: accents, switching to Russian mid-sentence, banking terminology.
The reason is data. Open Uzbek speech corpora exist, but they are read speech, not contact-centre calls: the Uzbek Speech Corpus holds 105 hours from 958 speakers, and its authors report a 17.4% word error rate on the test set (Musaev et al., 2021); FeruzaSpeech adds 60 hours from a single speaker, with transcripts in both Cyrillic and Latin script (Povey, 2024). Neither contains phone conversations — that data a team has to collect and label itself.
What works out of the box
The base model handles clean studio speech and standard phrases well. For FAQ scenarios — “card balance”, “opening hours”, “branch address” — that is nearly enough. The trouble starts on the phone channel: compressed audio, noise, interruptions.
What we had to push
We raised accuracy on domain vocabulary not by swapping the model but with data: labelled real calls, a glossary of terms, handling of “Uzbek ↔ Russian” code-switching. Latency was a separate fight — streaming recognition instead of waiting for the end of a phrase removes the most noticeable pause.
In voice AI it is not the largest model that wins, but the shortest, most predictable path from sound to answer.
| Pipeline stage | Base model out of the box | What we had to push |
|---|---|---|
| Clean studio speech | Recognised reliably | Nothing — the base model is enough |
| Phone channel | Accuracy drops: compressed audio, noise, interruptions | Fine-tuning on labelled real calls |
| Domain vocabulary (bank, logistics, clinic) | Terms get confused and flattened into common words | A glossary of terms and labelled industry dialogues |
| “Uzbek ↔ Russian” switching | The phrase is lost at the language seam | Dedicated code-switching handling |
| Latency before the answer | Waiting for the end of a phrase creates an audible pause | Streaming recognition instead of batch |
What breaks Uzbek speech to text in production
The table lists five reasons why recognition that is confident on a clean recording makes mistakes on a live call: how each one shows up and what is done about it.
| What breaks it | How it shows up | What is done about it |
|---|---|---|
| Uzbek ↔ Russian switching inside a sentence | The part after the language seam is lost — which is where amounts and service names are spoken | Dedicated code-switching handling, mixed-speech examples in the labelled data |
| Dialects and accents | Regional words are replaced with similar standard ones | Labelling calls from different regions, testing on the customer’s own recordings |
| 8 kHz phone channel | Compressed audio, noise and interruptions: a model that is confident on a clean recording drops words | Fine-tuning on real phone recordings |
| Proper names and amounts | Surnames, service names and numbers are written down wrong, terms are flattened into common words | A glossary of terms and names, labelled industry dialogues |
| Latin and Cyrillic script in the output | The same text arrives in different scripts, so search and reports stop matching | One script for transcripts, normalising the text to a single alphabet |
How the same task looks from the business side — what ends up in the transcript and how to connect call recording — is covered in the article on transcribing calls in Uzbek. Zukko.AI runs its own speech recognition models for Uzbek, Russian and English; the voice AI agent and call analytics are built on them.
Takeaway for anyone choosing a voice vendor
Ask not “which model do you use” but “on what data is it fine-tuned for our language and our industry” and “what is the latency on the phone channel”. For CIS banks those are the two decisive questions. Exact figures for your environment are confirmed on a demo. How this stack works inside the platform is shown on the Zukko.AI Uzbek speech to text page.
Frequently asked questions
Why is Uzbek speech recognition harder than Russian
There is less training data and the speech is alive: strong accents, dialect variation and switching to Russian mid-sentence. Add a phone channel with compressed audio and noise, and a generic model that sounds great on a studio recording starts dropping words. That is why quality is decided by labelled real calls rather than by model parameters.
What do you gain by moving to a larger model
Less than people expect. A larger model lifts general quality but still does not know your industry vocabulary, and it adds latency. Accuracy on domain terms comes from fine-tuning on labelled real calls and a glossary, not from model size on its own. So check the data and the channel first, and argue about model size afterwards.
What latency is acceptable in a voice conversation
The benchmark is simple: the person must not feel a pause after finishing a sentence. So you measure the whole path — recognition, understanding, response and synthesis — not just model runtime. Streaming recognition removes the most noticeable part of the wait, because the system starts processing the phrase while the person is still speaking.
How does the AI handle mixed Uzbek-Russian speech
Code-switching is handled separately: the system does not force the speaker to stay in one language, it detects the seam and continues the phrase. Without that handling a model loses the second half of the sentence — which is usually where numbers, amounts and service names appear. In Uzbekistan this is ordinary speech, not an edge case.
What should you ask a voice AI vendor before launch
Three things: on what data the model is fine-tuned for your language and industry, what the latency is specifically on the phone channel, and what happens on an out-of-scope request. Ask for a test on your recordings — exact figures are confirmed on a demo, not in a deck. Also ask which script the transcript arrives in: Latin or Cyrillic.
Do you need your own models or are off-the-shelf ones enough
It depends on the segment. Small and mid-size businesses usually need only ready models tuned to their catalogue; launch takes from 2 hours. For enterprise with data and environment requirements the speech stack is deployed separately, and the scope is agreed during rollout — up to 30 days. Zukko.AI runs its own speech recognition models for Uzbek, Russian and English.
Is there open data for Uzbek speech to text
Yes, but not much, and it is read speech. The open USC corpus holds 105 hours from 958 speakers (2021), and FeruzaSpeech holds 60 hours from a single speaker with transcripts in Cyrillic and Latin script (2024). Neither contains contact-centre phone conversations, so teams collect and label data for their own channel and industry themselves.