Speech from PowerShell

Future planning research only. This document does not describe a working speech build or authorize any change to stock Kokoro semantics. The direct PowerShell phoneme-to-PCM model and physical speaker gate take precedence.

SMA may own text planning. A language model may supply authored text, communicative objective, discourse annotations, or token surprisal, but it does not write the phoneme stream and it does not mix stage directions into the transcript.

authored text + objective + optional model evidence
    -> coordinate-preserving SMA parse
    -> speech-act and boundary candidates
    -> spoken normalization
    -> admitted phoneme adapter (checked against a pinned oracle)
    -> Kokoro inputs
    -> device acoustic and listening gates

speaker is a stable discourse identity; voice is the selected Kokoro style asset. They are separate on purpose: a character can be recast without changing the dialogue plan, and one voice can serve multiple anonymous roles.

Representations

• Authored text is immutable and round-trips byte-for-byte through the SMA lexer.
• Cue cards use ⟦name[:parameter]⟧. They are explicit overrides and debug fixtures, not the normal authoring language. They never enter visible or spoken text.
• Speech plan records source coordinates, speech-act candidates, planning load, recharge decisions, and optional surprisal evidence.
• Spoken text contains only material intended for pronunciation.
• Phonemes and PCM are later derived artifacts with their own admission gates.

Multiple speakers

Every speech plan and phrase bundle carries both speaker and voice. A speaker turn selects a verified voice/style resource without reloading the model DLL. The intended pipeline keeps the model session and one bounded AAudio stream alive across phrases. Device results name both fields so a listening result remains attributable; this multi-speaker path is not yet device-proven.

This is a target, not evidence-free: each admitted voice asset must be pinned by hash, each phrase still passes the FP32 oracle gate, and the first multi-speaker claim requires a physical speaker result. Voice resources belong in the verified model DLL, not a host export step at inference time.

The initial cue vocabulary is deliberately small: breath, pause, laugh, cough, clear_throat, sigh, and gasp. Pauses require an explicit duration from 20 to 5000 milliseconds. breath denotes a planning boundary to be rendered as a measured short pause, not an inhalation sound. Other nonverbal cues are not proven audio outputs. Asterisks and ordinary square brackets have no special meaning.

Borrowed mechanisms

Two local research repositories provide useful, bounded mechanisms:

• ChangeModel separates canonical state from represented state, records predicted and actual deltas, and promotes a representation mutation only when full-ledger replay strictly reduces contradictions or error. Speech planning can use the same mechanism to discover whether features such as question type, discourse contrast, surprisal, remaining breath budget, or dialogue role explain measured acoustic outcomes.
• JS2PS preserves original coordinates, proposes small reversible edits, scores parser shape separately from semantics, and admits a result only through an execution oracle. Speech lowering follows the same discipline: an inferred contour is a candidate, and a punctuation, phoneme, duration, or style change is promoted only after reference, device, and listening gates.

Neither repository proves a speech model. Their search and admission patterns are the reusable parts.

Evidence admitted so far

Primary studies support the architecture, but not a universal fixed syllable limit:

• Wang et al. measured breathing and audio from 16 healthy North American English speakers. Inspiratory locations usually aligned with grammatical boundaries; reading and spontaneous speech differed, with mean breath-group durations of 4.05 and 4.88 seconds in their data. This supports a dynamic state tied to task and syntax rather than a hard “16 syllables” rule. Primary paper
• Rochet-Capellan and Fuchs measured respiratory kinematics in 26 speakers of spontaneous German. Breath-group duration averaged about 3.5 seconds, and inhalation depth and duration varied with upcoming clause type and group length. This supports planning a recharge from both remaining state and upcoming linguistic load. Primary paper
• Nickerson and Chu-Carroll elicited direct and indirect readings of identical yes/no questions in dialogue. Boundary tone and final pitch differed with intended speech act, showing why punctuation alone is insufficient. Primary paper
• Chodroff and Cole manipulated information structure and affect in American English mini-stories. Contrastive focus changed F0 excursion, while lively affect also changed duration and amplitude. This supports treating model surprisal and discourse focus as prominence evidence rather than literal synthesis commands. Primary paper
• Karlapati et al. used contextual text and parse-tree features to sample prosodic representations for neural TTS, with controlled listening tests. Their results support context-conditioned prosody while also leaving a gap between predicted and oracle prosody. Primary paper

These sources justify candidate features and test design. They do not establish that Kokoro exposes the corresponding controls or that any proposed mutation sounds better on this device.

Admission sequence

1. Preserve authored text and reject malformed cue cards.
2. Produce speech-act, boundary, state, and prominence candidates without changing authored text.
3. Lower one candidate through a reversible mutation.
4. Verify pronunciation and tensor bounds.
5. Measure the physical device and play the result through the speaker.
6. Compare blinded or paired listening judgments and retain the full outcome ledger, including rejected candidates.

The current implementation proves only the first two steps. Contour labels and word-count planning budgets are explicitly provisional until later gates pass.