What is ASR: speech recognition in plain words, why phone calls are hard and its role in an AI operator
Sado AI team · · 4 min read
In short
ASR (automatic speech recognition) is technology that turns spoken words into text. It’s also called STT (speech-to-text). On a phone call, ASR hears the caller and writes it down: “tomorrow at seven in the evening, a table for six”. ASR doesn’t answer and doesn’t understand meaning on its own — it’s only the ears. In an AI voice operator, ASR is the first link of the chain: speech becomes text, the operator works out the meaning, builds an answer and says it aloud.
How ASR works
- The audio is cut into short slices, and the sound in each slice is identified.
- A model matches the sounds to words and produces several candidates.
- A part that knows the language picks the one that makes sense: in Uzbek, “olti kishiga” (for six people) rather than the similar-sounding “oltin kishiga” (for golden people).
- The result is text, a transcript. Many systems also mark when each word was said.
Quality is often measured as WER (word error rate): how many words out of 100 are wrong. For a business another question matters more: were the name, number, date and address understood correctly?
Example: a restaurant booking in Yakkasaray
Bahor restaurant (name invented) gets many booking calls in the evening. ASR might write down what callers say like this:
| Caller said | ASR wrote | What to do |
|---|---|---|
| “Tomorrow at seven, a table for six” (in Uzbek) | the same, word for word | All correct, book it |
| “Banket zalingiz bormi, svadbaga” (Uzbek with a Russian word) | the same, word for word | Mixed speech, the call stays in Uzbek |
| From the street: “Chilonzordan kelamiz” | “shilondan kelamiz” | Noise: ask for the word again |
| “My number is 90 123, 45, 67” | “my number is 90 123 45” | Number incomplete: read it all back and confirm |
Mistakes are normal; what matters is what happens next. A good operator, human or AI, reads the name, number and time back and asks the caller to confirm.
What makes ASR harder on the phone
- Narrow phone audio: low and high tones are cut off, so “s”, “sh” and “ch” sound alike.
- Noise: the street, a car, a bazaar, children at home.
- Mixed speech: Uzbek and Russian words in one sentence.
- Dialects and fast speech.
- Names, street and product names, numbers and dates.
- Two people talking at once, or the caller turning to someone beside them.
ASR’s role in an AI operator
ASR only listens. Answering a call takes the whole chain: recognising speech, understanding it, building an answer from the company’s information, saying it in a natural voice (what is TTS), connecting to the phone line and writing the result to the CRM. So a business searching for “Uzbek speech recognition” usually needs a ready operator rather than a separate service.
Sado, an Uzbek-speaking AI operator, understands Uzbek and Russian. When Russian words slip into Uzbek speech, the call stays in Uzbek, and it switches language when the caller asks. The recording and full transcript of every call stay in the portal, so you see what the operator heard and how it answered. More: an Uzbek and Russian voice operator.
Testing ASR: a checklist
- Call from a noisy place, such as the street.
- Dictate a phone number in groups: “90, 123, 45, 67”.
- Say a date and time in words: “the fifteenth of November, six in the evening”.
- Say a first name, a surname and a street name.
- Mix Russian words into an Uzbek sentence, then switch to Russian.
- After the call, read the transcript and compare it with the recording.
Test it yourself: call 78 778 01 23 or try it live. About the AI operator for incoming calls and other questions: questions about Sado.
Questions
Are ASR and STT the same?
Yes. ASR is automatic speech recognition, STT is speech-to-text. Both name the same technology that turns a voice into text.
Does ASR understand Uzbek?
Systems that recognise Uzbek exist, but quality varies, especially on phone audio and mixed speech. Test in the language your customers speak, and from a noisy place.
Does Sado keep call transcripts?
Yes, the recording and full transcript of every call stay in the portal, whatever the language. Data is protected: storage time is limited and you can ask for deletion.