Build this as two steps: convert speech to text, then use that text to suggest values for the form. Reading text aloud is a different feature and does not prove that transcription is available.
A browser-based option uses speech recognition where the browser supports it. MDN's speech recognition guide documents limited browser availability and notes that some implementations send audio to a recognition service. Do not promise that this works offline or keeps audio entirely on the device. Keep a typed-input fallback and test the actual phones and languages your users need.
If you choose an outside transcription provider, verify its supported audio format, limits, price and retention terms before connecting it through server-side code. Do not assume that the app's built-in voice narration endpoint accepts speech as input.
After transcription, ask the builder to extract only the fields your form supports and show the proposed values for review. Validate them on the server and save only after the person confirms. A mistaken name, date or amount should be easy to correct.
Built-in AI extraction uses workspace credits; an outside transcription provider can charge separately. Decide explicitly whether the app stores recordings or transcripts, and inspect the resulting data flow. Omitting an audio table does not establish the provider's retention policy.