Voice pipeline switches

The voice pipeline has a few switches, set in the portal's environment, that turn its optimizations off, so that what each one buys can be measured with the timing profiler: run a second portal beside the first, with the same model and voice, and change one switch at a time. None of them is needed in normal use. Left alone, a portal runs the full pipeline.

Progressive transcription (the recording is transcribed while you still speak) and streamed playback have no switches of their own. They belong to the default parallel mode, and VOICE_PIPELINE_MODE=sequential turns both off together. VOICE_SENTENCE_CHUNKS and VOICE_TTS_PREFETCH are the two steps back up from the sequential baseline, and have no effect unless the mode is sequential.

Variable	Default	Effect
VOICE_PIPELINE_MODE	parallel	sequential is the baseline: speech is transcribed only after it has finished, the reply goes to speech synthesis only once the whole agent turn is done, and every generated chunk is buffered completely before playback starts. Thinking and compaction filler speech is suppressed so it cannot overlap the measured work. Cancelling and muting still work.
VOICE_SENTENCE_CHUNKS	false	Only in the sequential mode. With true, sentences are submitted to speech synthesis during generation, and each sentence's audio plays before the next is synthesized. Audio is still fully buffered per sentence. The first step up from the baseline.
VOICE_TTS_PREFETCH	false	Only in the sequential mode. With true, one synthesis producer runs alongside ordered playback, with at most two prepared phrases queued ahead. Each sentence's audio is still buffered before it plays.
VOICE_SKIP_FIRST_THINKING	true	false keeps the model's normal thinking on the first response of a voice turn, which is otherwise switched off, for a local model, to start speaking sooner.
VOICE_RESPONSE_INSTRUCTIONS	true	false sends no speaking instructions: no [Audio mode] marker on the input and no voice rule in the system prompt, so replies are not asked to be brief, plain text or canvas-first. See Speaking instructions.
VOICE_STATUS_SPEECH	true	false keeps the spoken status lines out: the thinking and compaction feedback.
VOICE_COMPARISON	false	true shows "Streaming pipeline" beside the voice session's title, and gives the login a cookie of its own (see below).

The voice screen's title says which stage a portal is on. In the sequential mode it reads "Sequential baseline", "Sentence chunks · buffered audio" or "Sentence pipeline · buffered audio", depending on the two sentence switches; with VOICE_COMPARISON=true outside it, "Streaming pipeline".

In the sequential mode, and with VOICE_COMPARISON=true, the login cookie is pi_portal_sequential_auth instead of pi_portal_auth, so a comparison portal running beside the main one on the same host keeps its own login (browsers share cookies between ports). Changing the mode changes the cookie, so log in again afterwards.

Recording a comparison

• Use the same prompt, model, speaking voice, guidance mode, speech-detection settings and starting context in both. Start a new session for each take; an earlier conversation still influences the model.
• Record one portal at a time when both share a model server and speech services, and end voice mode on the other one first.
• Record a warm-up apart from the measured turns, enable the profiler, and export its report after each take.
• Go up one step at a time: the sequential baseline; VOICE_SENTENCE_CHUNKS=true added; VOICE_TTS_PREFETCH=true added to that; then the default parallel mode, which brings progressive transcription and streamed playback together.

The metric is the time from the last detected speech to the estimated start of the reply's audio. Browser output timing is a software estimate, not an acoustic measurement. Progressive transcription often finishes before the end of speech is detected, but a final request is still made when what it has does not cover the last of the speech, and streamed playback keeps its startup cushion: do not describe either as zero work or zero buffering.