Speech-to-text crossed the threshold from research problem to utility. That is worth being precise about: accurate transcription is now assumed, which means it is no longer where anything differentiating happens. The interesting questions are what you do with the text, and when voice is the right input at all.
The pipeline, and where the value sits
- Capture — quality here bounds everything downstream. Far-field audio with three people talking over each other is a hardware problem no model fixes.
- Transcription — a commodity, and reliable enough to stop thinking about for most languages.
- Diarisation — who spoke. Harder than transcription, and usually where accuracy actually breaks.
- Understanding — decisions, commitments, questions, sentiment. Where the value is.
- Action — what the system does with it. Where the value is realised or lost.
Where voice is genuinely the right interface
- Hands and eyes busy. Field work, clinical settings, driving, warehouses — voice is not a preference, it is the only free channel.
- Faster than typing on the device available. Dictation on a phone beats a thumb keyboard for anything longer than a sentence.
- The conversation is the artefact. Meetings, interviews, support calls — the audio exists whether or not you use it.
- Accessibility. For some users, voice is the difference between usable and not.
Where it is the wrong interface
- Anything requiring precision the speaker cannot verify — long identifiers, code, exact figures.
- Environments where speaking aloud is socially or physically impractical, which is most open-plan offices.
- Interactions needing a visible, scannable list of options. Speech is linear; a menu is not.
- Anything the user will want to review and edit before committing — unless you present the text for editing, at which point the keyboard is back.
The failure that repeats across voice products
Building the transcript as the deliverable. A searchable wall of speech is barely more useful than the recording — the moment someone decided something is still buried forty minutes in. Extract decisions, commitments and open questions as structured records; keep the transcript as provenance, not as the product.
What breaks in production
- Accents, domain vocabulary and code-switching. Benchmark accuracy is measured on clean speech; your users say your product's names, which are not in any training set. Custom vocabulary helps materially.
- Overlapping speech. Diarisation degrades sharply, and meetings are mostly overlapping speech.
- Silence and filler. Raw transcripts are full of both, and extraction quality improves noticeably after cleaning.
- Latency expectations. Real-time and batch are different products with different architectures; retrofitting one into the other rarely goes well.
Consent is a product requirement
Recording law varies by jurisdiction and two-party consent is not universal. Beyond the legal position, recording changes what people say — which is a product consideration as much as a compliance one. Announce it, make opting out genuinely possible, and set retention deliberately rather than defaulting to forever.
Frequently Asked Questions
How accurate does transcription need to be?
Less than people assume for extraction, which works on meaning, and much more for any verbatim quote you surface as evidence. Set the bar per use, not globally.
Should audio be stored after transcription?
Only with a reason. Audio is the highest-sensitivity artefact in the pipeline and the most expensive to retain. If you keep it for provenance, scope retention to that purpose.
Is on-device transcription viable?
Increasingly, and it removes the subprocessor question entirely - which in regulated settings is often worth more than the accuracy difference.
References
About Jishu Labs
Jishu Labs is a software development company founded in 2016. We build custom software, AI/ML systems, and full-stack web and mobile applications for clients, and we make eight AI tools for software teams.