Punjabi speech · transparent evidence
Test the audio.
See what the model heard.
Transcription tests run without playback. Listen only if you deliberately use a separate audio control. The reference, errors, route-critical words, latency, request ID, and estimated API cost stay visible together.
Installed evidence pack
One-click silent transcription tests
Loading the local sample manifest…
Your voice · private evaluation only
Record Punjabi or choose an audio file
Recording stays in this browser until you press Transcribe my clip. The navigator does not persist raw audio or transcripts. Do not say a name, passport number, UCI, or exact address.
No recording selected.
Add a reference to calculate quality scores
Native-speaker critical route prompts
A native evaluator can open the 12-scenario manual seed pack and express each brief naturally in Punjabi to test ASR in the private lab. A separate automated fixture passes each prestructured request to the catalog matcher and asserts its declared route. These are not an end-to-end speech-to-route result; the 200-case launch set must be independently authored, held out, and run through the full flow.
Run evidence
Results
No model calls yet. Results appear here after you choose a test.
How to read these numbers
WER is word error rate. CER is character error rate. Lower is better; either can exceed 100%.
Critical words checks whether pre-declared locations, needs, or numbers survived transcription.
If the transcription endpoint does not return a route, route readiness is only a proxy: it means every pre-declared critical word was retained. It is not a live routing verdict.
Cost uses the configured transcription price and clip duration; billing records remain authoritative.