Building voice assistants for services in Uzbekistan
A voice interface earns its place where the customer will not or cannot type. But the current quality of Uzbek speech recognition sets hard limits on what is feasible.
Where voice genuinely fits
Realistic scenarios on this market: taking calls outside business hours, automatic routing in a call centre, order confirmations, appointment reminders, and the voice messages many users send in Telegram instead of typing.
That last one is underrated. A substantial part of the audience, particularly older users and those in the regions, sends the bot voice notes instead of text. A bot that cannot process them simply does not work for those people.
Poor fits: long consultations, discussing complex terms, anything requiring precision with numbers and names. Voice is bad at conveying an address or a card number, and automating that ends in repeated clarifications.
Uzbek speech recognition
Recognition quality for Russian and English is high; for Uzbek it is noticeably lower, especially with conversational speech containing Russian borrowings — which is the norm in everyday Tashkent speech.
Regional dialect differences add difficulty: Khorezm and Fergana pronunciation are recognized with differing accuracy, and a model trained on the literary norm will make mistakes.
The practical conclusion: do not build critical flows on full comprehension of free-form Uzbek speech. Constrain the dialogue to short predictable answers and always confirm what was recognized before acting.
Architecture and latency
The chain is standard: speech recognition, model processing, speech synthesis. The problem is cumulative latency: three links of a second each produce a pause a person on a phone call reads as a dropped connection.
Techniques: streaming recognition instead of waiting for the end of the utterance, short acknowledgement phrases during processing, and pre-synthesized audio for standard lines instead of synthesizing on the fly.
For telephony, account for channel quality. Narrowband audio on a mobile network cuts frequencies and degrades recognition accuracy considerably more than a messenger microphone recording.
Speech synthesis
Uzbek speech synthesis sounds less natural than Russian: intonation suffers and stress in proper nouns and place names is frequently misplaced.
For frequently repeated phrases use audio recorded by a voice actor and reserve synthesis for the variable parts. The hybrid approach sounds noticeably better at the same cost.
Pronounce numbers and amounts through verified rules rather than leaving it to the synthesizer. A sum in UZS read out incorrectly devalues the entire conversation.
Expectations and cost
Market reference: a voice bot handling calls with a simple flow runs 25M – 50M UZS. With full CRM and telephony integration the figure is higher and depends heavily on the PBX in use.
Running costs combine recognition minutes, model tokens and synthesis minutes. Voice is appreciably more expensive than text per conversation, and that belongs in the payback calculation.
A realistic expectation: a voice bot closes simple routine contacts and relieves the first line, but does not replace an agent in complex cases. Promises of fully replacing a call centre at the current state of Uzbek-language technology do not match reality.