Skip to main content

voice

Since v2.37.0 · Top-level declaration

Grammar​

voice <Name> {
    stt: <Upstream|Preset@vN>        # cascaded: BOTH stt: and tts:
    tts: <Upstream|Preset@vN>
    realtime: <Upstream|Preset@vN>   # fused: exactly one (XOR stt/tts)
    carrier: mulaw8k | pcm16         # default mulaw8k (PSTN)
    interruptible: true | false # barge-in via v2.36.0 interrupt/resume
    legal_basis: <basis>             # REQUIRED when interruptible
    persona: <PersonaName>           # optional
    context: <ContextName>           # optional
}

voice is the simplicity layer over v2.37.0's upstream: a working, swappable, auditable phone agent in under 20 lines for any blessed vendor preset — without giving up a single guarantee, because voice is pure macro-expansion to the primitives already in the language.

voice Concierge {
    stt: DeepgramSTT@v1
    tts: ElevenLabsTTS@v1
    interruptible: true
    legal_basis: legitimate_interest
}

That declaration expands to — and axon desugar prints — exactly:

  • the ots μ-law↔PCM16 codec pair (carrier mulaw8k, the PSTN default; pcm16 skips it),
  • a carrier-facing session ConciergeCarrierTurn + socket ConciergeCall with backpressure: credit(8) — and, because interruptible: true, a v2.36.0 interrupt/resume region for barge-in plus reconnect: cognitive_state and the declared legal_basis (the parked mid-utterance residual is a governed data-at-rest surface),
  • one upstream per vendor leg (ConciergeSttLink from DeepgramSTT@v1, ConciergeTtsLink from ElevenLabsTTS@v1), themselves expanded from the v2.37.0 preset catalog.

One grammar, both architectures​

Cascaded (stt: + tts:) and fused speech-to-speech (realtime:) are equally first-class — the same voice grammar, never a special-cased second path:

voice Live {
    realtime: OpenAIRealtime@v1
    carrier: pcm16
}

The laws (axon-T852)​

  • stt:+tts: XOR realtime: — cascaded needs both legs, fused exactly one.
  • interruptible: true requires legal_basis: — the sugar cannot generate a program the ParkedResidualSoundness proof refutes.
  • Every leg must resolve: a Preset@vN from the catalog, or a declared upstream (which is then referenced, never re-declared).
  • The expansion earns no exemption: the generated session/socket/ upstreams are checked by the ordinary v2.3.0/v2.36.0/v2.37.0 laws and carry the ordinary PCC proofs (UpstreamProjectionSoundness, InterruptibleSessionSoundness, ParkedResidualSoundness).

Never a black box​

voice itself never reaches the IR — the deployed artifact IS the expansion, and axon desugar <file> prints it verbatim. A simplicity layer a compliance reviewer cannot see through would undo the audit-by-construction property; this one is the seeing-through. Swapping Deepgram for AssemblyAI is a one-token edit; the seventh vendor the market ships next quarter is a hand-written upstream declaration away, not a rewrite.