Audio streaming

Audio breaks into two subtypes that map to TAP differently:

1. Generated audio (text-to-speech, on-the-fly music synthesis) — this is structurally identical to LLM inference. The producer generates output as the consumer is paying for it; halt-on-quality makes sense; per-second metering aligns with the unit of value.
2. Catalog audio (music streaming, podcasts) — the platform-aggregator pattern from Video applies. Don't put the meter in front of the listener.

Generated audio (TTS, music synthesis)

The session shape is unchanged from the LLM mapping. What differs is the quality signal:

LLM session	Generated-audio session
output_token	output_second (or output_chunk)
repetition_guard	encoder-error / silence-detection evaluator
json_schema	format-conformance evaluator (sample rate, channels)
topic_drift	spoken-content drift (semantic check on transcript)

The producer's tokenizer_id becomes a chunk_unit_id declaring what the per-second unit means (e.g. pcm-s16le-22050). The consumer can sanity-check incoming chunks against this declaration and halt on encoding violation just as the LLM consumer halts on JSON violation.

A useful concrete evaluator for TTS:

def silence_guard(*, max_silent_seconds: int = 3) -> Evaluator:
    """Halt if more than max_silent_seconds of silence appear in the
    accumulated waveform. Catches the most common TTS failure mode —
    the model hung mid-sentence and is producing zeros."""
    # accumulated buffer; check trailing window for RMS below floor
    ...

Catalog audio (Spotify-shaped products)

Same lessons as video:

• Listeners should not see a per-second meter.
• The platform plays consumer; the producer (artist, podcast host) plays producer.
• Settle in a platform-issued token whose redemption value is a function of monthly revenue.

The streaming math is even simpler than video — there is no quality halt to speak of in catalog audio. The platform's per-second commitment is the entire mechanism: it accrues during play, signs periodic commits, settles at month-end. Producers see exactly how many seconds of their work were played, with on-chain receipts.

A specific product idea: split-track streaming

One area where per-second TAP commitments can be exposed to end-users without breaking UX is interactive remixing — apps where a listener loads multiple tracks and pays per-second only for the stems actually audible at any moment. The cognitive overhead of "is this stem worth paying for" is already part of the creative interaction; the meter is a feature, not friction. Per-stem channels with bilateral halt straightforwardly enable this.