AI search

AI voice search: talk to the catalogue, hands free

Voice search is not a novelty on a South African storefront. It is an accessibility feature, a mobile-data feature and a trade-counter feature: someone with a trolley in one hand and a phone in the other will speak a query they would never type.

GoGee's voice search does one job cleanly. It captures audio, transcribes it, and hands the transcript to the same AI search pipeline that powers typed search, so voice inherits filter extraction, semantic matching and exact-SKU scoring for free.

GoGee feature series · 3 of 33

How it's actually built

Transcription model: openai/gpt-4o-mini-transcribe
Accepted audio: WebM, MP4, MPEG, M4A, AAC, signature checked
Size cap: 2MB per clip
Rate limits: 15/5min and 40/hour per visitor, 600/hour site-wide
Fallback: Browser-native speech recognition
Downstream: Transcript enters the standard AI search pipeline

Voice pipeline

01Input
Browser recording
Clip captured client-side
02Deterministic
Validate audio
Type normalised, EBML or ftyp signature verified, 2MB cap
03AI
Transcribe
openai/gpt-4o-mini-transcribe returns a plain transcript
04Human
Shopper confirms
Transcript shown before results load so people self-correct
05Output
Standard AI search
Transcript enters the text pipeline

Guardrails and fallbacks

Browser-native speech recognition fallback
15/5min and 40/hour per visitor
600/hour site-wide

AI voice search, data flow, generated from the shared GoGee feature diagram template.

Capture, validate, transcribe

The clip is recorded in the browser and posted to a server function. Before any paid transcription happens, the audio type is normalised, the file signature is verified, an EBML header for WebM, an ftyp box for MP4, and the size is capped at 2MB. Rate limits apply per visitor and site-wide.

Only then is the audio sent for transcription. The response is a plain transcript, which the interface shows the shopper so they can see what was heard before results load. When people can see the transcription, they self-correct instead of blaming your site.

Graceful degradation is the whole design

If transcription is unavailable for any reason, a provider error, a missing key, a network failure, the server returns an empty transcript flagged as unavailable, and the browser's own speech recognition takes over. The mic button never becomes a dead button.

This is the same philosophy as the rest of GoGee's AI: every AI path has a deterministic path behind it, because a shopper does not care which layer failed.

Voice reuses the search brain

Once transcribed, "do you have those big blue crates for under two hundred rand" is exactly the same input as if it were typed. Filters are extracted, the phrase is embedded, exact identifiers are scored above similarity, and stock and branch context still apply.

The search log records the input mode, so voice, photo and typed searches can be compared. In practice that reporting is how clients decide whether to promote the mic button on mobile, and on catalogue-heavy stores, they usually do.

Questions we get asked

Which languages does voice search handle?

The transcription model handles multilingual speech, and search then runs on the resulting text. For on-page translation across the official South African languages, that is a separate feature, the AI language switcher.

Is the audio stored?

The clip is sent for transcription and the search is logged as a voice-mode search with its transcript. Audio is not needed after the transcript exists.