How It Works¶
Last updated: 2026-09-16
Kawa is a composable transcription pipeline. The first managed API, available now, handles post-conversation workloads: recorded calls, voicemail, and large-scale analysis of archived customer interactions.
A real-time API for live captioning, voice agents, and in-conversation analytics is planned; see the roadmap.
Both run the same five stages, each replaceable so you can tune for your language, domain, or latency budget.
The five stages¶
- Audio resampling
- Voice activity detection with speaker diarisation
- Noise filtering and audio enhancement
- Speech-to-text inference
- Post-processing (normalisation, temporal ordering, speaker labelling, structured output)
Summaries and Tags¶
Once a transcription job has succeeded, it can be sent through two further, on-demand steps: summarisation, which produces a topic and prose summary, and tagging, which classifies the conversation against a fixed taxonomy (call reason, customer sentiment, and so on). Each is a separate job in its own right, created from an existing transcription via POST /v1/summarize or POST /v1/tag, and polled the same way. Unlike the five pipeline stages above, these are opt-in per transcription rather than run on every upload.
Either endpoint also accepts plain text directly (transcript_text), bypassing transcription and the audio pipeline entirely.
Terms used in this documentation¶
| Term | Meaning |
|---|---|
| Job | One submitted audio file, or one summary/tag request, tracked through queued, processing, succeeded, or failed. |
| Segment | One span of transcript text attributed to a single speaker, with start and end timestamps. |
| Speaker alias | A per-job label (User, Agent) assigned by diarisation. |
| Post-conversation | Transcription of a recording after the conversation has ended, as opposed to real-time, in-call transcription. |
| Job type | transcription, summary, or tags — which of the three result shapes a job's result carries. |
Data Residency¶
Kawa is hosted in Singapore and keeps customer audio and transcripts in-country, supporting organisations under Singapore's Personal Data Protection Act (PDPA) and sector frameworks including Monetary Authority of Singapore (MAS) guidelines.