A transcript is text, not a business decision
STT converts audio into text. “Tomorrow at half past three” may be transcribed correctly, but the exact date, time-of-day context and availability need separate interpretation and validation. Recognition success is not permission to create a booking.
Keep verbatim text separate from normalized business fields. This helps distinguish a recognition error from an interpretation or workflow error. A model must not invent another person while “correcting” an uncertain name.
Separate partial and final transcription
Streaming recognition returns text while speech arrives. Partial results may change as later sounds clarify an earlier word. A final result is not infallible: it means recognition of that segment has finished.
Associate events with the relevant segment and sequence. A late partial transcript must not overwrite a corrected or confirmed value. That is a state-management responsibility, not a problem solved solely by selecting an STT model.
| Event | Verbatim text | Business interpretation |
|---|---|---|
| Partial | “ხვალ სამ…” — tomorrow at th… | Unconfirmed candidate time |
| Final | “ხვალ სამზე, არა, ოთხზე” — tomorrow at three, no, four | Interpretation must respect the correction |
| Context validation | Final transcript retained unchanged | 16:00 on the known date only with the corresponding daytime context |
| Stale partial event | “ხვალ სამზე” — tomorrow at three | Must not replace the newer time |
What needs special attention in Georgian calls?
Include inflected names, surnames, short confirmations, amounts, phone digits and Georgian time expressions. “ლუკასთან” refers to the name Luka in context. Georgian “ოთხის ნახევარი” means half past three; in an afternoon context that is 15:30, not 16:30.
Evaluate code-switching, brand names and unfamiliar surnames separately. Validate critical fields against their format, an authorized catalogue or a caller clarification. Similar pronunciation is not proof of identity.
- Compare digit sequences exactly, not just the general sentence meaning.
- Vocabulary hints may help, but actual support depends on the provider.
- A misrecognized short “yes” must not become an unconfirmed business action.
Telephone audio and clean microphones are different test conditions
An 8 kHz telephone recording differs from a clean microphone recording. Codec, compression, packet loss, echo and overlapping voices matter. These are test conditions, not measured accuracy results for a particular system.
Voice activity detection identifies speech activity; endpointing helps determine the end of a segment. Neither alone establishes whether speech comes from the caller or a television. Noise reduction can also damage useful speech, so evaluate background filtering and protection of the caller together.
How should recognition quality be compared?
Use consented recordings or a purpose-recorded test set with manually verified reference transcripts. Describe speaker count, conditions, accents, duration and critical-field coverage. Publish numerical results only after actually measuring them.
Word error rate (WER) counts substitutions, deletions and insertions relative to reference words. Character error rate (CER) applies the same idea to characters. A low overall error rate does not prove that a phone number or time is correct, so measure field and task accuracy separately.
| Measure | What to assess |
|---|---|
| WER / CER | Difference from a verified reference with a declared normalization policy |
| Critical fields | Exact name, date, time and phone-number recognition |
| Task outcome | Correct action or clarification after validation |
| Latency | Milliseconds from the end of speech to the final transcript |
| Uncertainty | Whether ambiguity triggers clarification instead of false certainty |
An integrated component is not a standalone product
This guide explains STT as part of a voice workflow. Integration does not establish OMO ownership of an underlying speech model or separate availability of a public transcription API, batch transcription or fine-tuning service.
Provider language support, access and limitations need verification for the implementation. One successful demo cannot replace representative telephone evaluation or guarantee recognition of every name and noisy utterance.
Common questions
Is STT the same as understanding?
No. STT produces text. Context, field meaning and action permission are evaluated elsewhere.
Are names always recognized correctly?
No. Unfamiliar names, inflections and similar sounds can cause errors. Clarify critical values or compare them with an authorized source.
Does a microphone demo prove telephone quality?
No. Codecs, 8 kHz audio, echo and poor connections require separate testing.
Does OMO own the recognition model?
Using or integrating a component does not establish ownership of its base model. Provider ownership and OMO’s engineering contribution should be assessed separately.
Can STT be purchased separately?
This page does not announce a standalone STT product or public API. Confirm any such request, scope and access terms with the team.
Primary sources and scope
Checked on September 17, 2026. These sources explain technical mechanisms, not the availability of every capability in an OMO implementation. Proposed connections and examples require a separate assessment.
- Google Cloud — streaming recognition
Interim and final results in streaming transcription.
- Google Cloud — measuring recognition accuracy
Reference transcripts and word error rate (WER).
- LiveKit — turn detection and interruptions
Distinct controls for turn-taking, speech activity and interruptions.