Default streaming STT prompt — deliberately empty: transcription is
unbiased unless an agent sets sttPrompt.
Worth knowing what an agent gives up by leaving it empty. Spoken identifiers
are the pipeline's quietest transcription failure: a caller spelling out a
confirmation code lands in an interim turn, and the formatted final turn
can revise those characters away entirely, so the code reaches the LLM
missing rather than misheard. The model still has a required tool argument to
fill, so it substitutes something plausible — in the worst case the example
value from the tool's own schema — and the turn fails with no error anywhere.
(Reproduced against FDB-v3: a spelled order ID was dropped from every final
turn, and the agent called the tool with the schema's example value.)
A prompt only helps when it is specific to the agent's own vocabulary, and
showing the spelled→joined form is what makes it stick, e.g.
"Callers read order IDs out character by character: 'K L 4 7 2' is KL472. Never omit a spoken identifier." A generic version of the same instruction
measured no better than none, which is why there is no default here — and
why an unrelated prompt is worse than empty: it biases the transcript
toward vocabulary the caller never used.
Default streaming STT
prompt— deliberately empty: transcription is unbiased unless an agent setssttPrompt.Worth knowing what an agent gives up by leaving it empty. Spoken identifiers are the pipeline's quietest transcription failure: a caller spelling out a confirmation code lands in an interim turn, and the formatted final turn can revise those characters away entirely, so the code reaches the LLM missing rather than misheard. The model still has a required tool argument to fill, so it substitutes something plausible — in the worst case the example value from the tool's own schema — and the turn fails with no error anywhere. (Reproduced against FDB-v3: a spelled order ID was dropped from every final turn, and the agent called the tool with the schema's example value.)
A prompt only helps when it is specific to the agent's own vocabulary, and showing the spelled→joined form is what makes it stick, e.g.
"Callers read order IDs out character by character: 'K L 4 7 2' is KL472. Never omit a spoken identifier."A generic version of the same instruction measured no better than none, which is why there is no default here — and why an unrelated prompt is worse than empty: it biases the transcript toward vocabulary the caller never used.