Longer Spanish phrases exposed the voice-transfer cost

Expand beyond the greeting

Six phrases per language were synthesized with the production Supertonic stack, converted, then scored with an independent Parakeet recognizer. The experiment used public demo narrator references and did not alter the production application.

The result that the greeting missed

English stayed at zero observed WER after four-step conversion. Spanish changed from 5.36% before conversion to 10.71% at four steps, 8.93% at eight and 7.14% at twelve. Some accent and proper-name differences already existed in the unconverted recognizer output.

Spanish median conversion time rose from 1.263 to 2.214 to 3.241 seconds. More steps reduced observed word errors slightly, at a substantial latency cost.

Cross-language transfer also completed six requests in each direction. Spanish speech using an English narrator reference reached cosine 0.813 and WER 10.71%; English using a Spanish reference reached 0.823 and zero observed WER.

What is ready

This establishes limited zero-shot transfer, not an exact voice clone or a seamless call. Keep Spanish conversion optional and experimental while testing content preservation, caching, streaming and actual two-device latency.

Evidence and publication boundary

This is an edited account of the recorded experiment. The linked archive is private; the methods, aggregate findings and limits are included above. Reference recordings, raw conversations, authentication material and workstation configuration are excluded.

• Source record 1