Making Punctuation Audible: A Boundary-Tone Plan for Sound Planning

Status: research / directional. The user's goal: it should be clearly audible when the text contains a question, exclamation, explicit period, dash, or ellipsis — each recognisably distinct from the others, not just from neutral text.

This note (1) surveys how modern TTS systems make these five marks audible, (2) measures exactly what VOICE_V8 does today and where the five marks collide or get dropped, and (3) proposes concrete sound-planning changes mapped onto dootdoot's existing primitives.

It does not change the voice contract. Every recommendation here is sample-affecting and would require a new voice version (V9) plus regenerated golden WAV hashes (CLAUDE.md "load-bearing invariants").

Summary

dootdoot already has the right machinery — per-mark final glides, boundary-strength pauses, declination/lowering, and a separate quiet-rest path for dash/ellipsis. But four of the five marks the user cares about are not reliably distinguishable today:

1. Period and exclamation share the same local boundary signature. Both map to a Falling final glide and the same 240 ms long pause (sequence.rs:360, synth.rs:83). They are not byte-identical end-to-end — ! raises the utterance arousal proxy, and a final ! gets the TerminalFlourish role — but the punctuation control itself still has no mark-specific punch. A period needs an audible settle; an exclamation needs an audible punch, especially away from the final syllable.
2. ... typed as three ASCII periods routes through period controls, not ellipsis. Only the single Unicode … (U+2026) is recognised as an ellipsis (sequence.rs:127). The far more common ... tokenises into three separate Period markers in --explain. The audio path does not literally render three falling glides or three additive 240 ms pauses because consecutive punctuation uses the first marker's glide and the longest single pause; still, the intended "half-finished thought" gesture almost never fires in practice.
3. Dash and ellipsis differ only by the length of a silent gap (340 ms vs 500 ms; sequence.rs:105,108). Both are "quiet rest + Hesitation role." Linguistically these are opposites — a dash is an abrupt cutoff, an ellipsis is a gradual trailing-off — and that contrast lives in the shape of the preceding syllable's tail, which today is identical for both.
4. The explicit period is weakly terminal. It has no dedicated role (it falls to the _ => ChattyReply arm of segment_role, performance.rs:414), and the clause marks (, ; :) all collapse to one neutral-glide Aside bucket. There is no continuation-vs-closure contrast to make a period read as "done."

The fix is not new primitives — it is assigning each of the five marks a distinct boundary-tone signature across the knobs we already have (final glide direction and magnitude, final lengthening, pause, brightness/intensity, and tail shape), plus a small normalisation pass so ... becomes an ellipsis.

There is also a source-of-truth mismatch to resolve before implementation: spec.md FR-82, docs/plan.md T-84, and design.md §8.8 say ASCII ... should already be a VOICE_V7 hesitation marker, but the current code and design.md §3.3 only treat the single Unicode ellipsis (…) that way. This note describes current VOICE_V8 behavior as observed from code and --explain; the V9 work should explicitly reconcile those documents.

1. How modern TTS makes these marks audible

Modern TTS — from ToBI-driven concatenative engines to SSML front-ends on neural voices — encodes terminal punctuation as a small set of boundary tones plus phrase-final timing and intensity adjustments. The canonical correlates:

Mark	Boundary tone (ToBI)	Pitch	Timing	Intensity	Percept
Statement .	L-L% (low fall to floor)	final fall to bottom of range, full declination	strong final lengthening, then settle	drops off	closure / "done"
Question ?	L* H-H% / H-H% (high rise)	final rise, declination suppressed	mild lengthening	sustained	open / "your turn"
Exclamation !	H* L-L% from a raised peak	expanded pitch range, higher peak, steeper move	sharper attack, can be faster	raised loudness / emphasis	emphatic punch
Clause , ; :	L-H% / H- (continuation rise)	slight rise or level, not to floor	short pause, light lengthening	sustained	"more coming"
Dash —	truncation / suspension	contour clipped mid-move, often a held or glottal stop	abrupt cutoff, then pause	cut short	interruption / break
Ellipsis …	half-cadent / level (L- or sustained)	level or shallow fall, held	lengthen + amplitude decay, long pause	fades out	trailing off / unfinished

Two themes matter for dootdoot:

• Direction and degree. A period and an exclamation can both end in a fall, but the exclamation falls from a raised, widened register with more loudness, and the period falls all the way to the declination floor with final lengthening. The contrast is in the magnitude and the surrounding range, not just the sign of the glide. SSML exposes this as separate pitch, range, rate, and volume/emphasis levers (SSML, milvus: prosody control).
• Dash ≠ ellipsis is a shape contrast. Both are pauses, but the dash is an abrupt break (truncated phonation, a suspended/held tone) while the ellipsis trails off (sustained level or "semi-cadent" pitch with amplitude decay and final lengthening — the classic "half-finished thought"). The two marks were historically the same glyph precisely because both mark a break; speech keeps them distinct through the tail of the preceding syllable, not the gap alone (Ellipsis, Boundary tone, ToBI).

2. What VOICE_V8 does today

2.1 The two punctuation paths

dootdoot splits punctuation across two enums:

• ProsodicPunctuation (sequence.rs:203) — ? . ! , ; :. Each contributes a final glide (sequence.rs:360) and a boundary pause (sequence.rs:353), and feeds segment_role (performance.rs:398) + the phrase-final prosody in phrase.rs (final_lowering_semitones, pitch_reset_semitones, pre_boundary_lengthening).
• HesitationMarker (sequence.rs:115) — - -- — – and …. These are not voiced; they impose a quiet, bridge-suppressed rest on the preceding syllable (sequence.rs:142) and push that segment into the Hesitation role.

2.2 Current per-mark signature

Mark	Parsed as	Final glide	Pause	Terminal role	Final lowering
?	Question	Rising	240 ms	TerminalFlourish (yelp/whistle)	suppressed (0)
!	Exclamation	Falling	240 ms	TerminalFlourish	sentence (−0.90 st)
.	Period	Falling	240 ms	default (ChattyReply)	sentence (−0.90 st)
,	Comma	Neutral	150 ms	Aside	clause (−0.20 st)
;	Semicolon	Neutral	150 ms	Aside	clause (−0.20 st)
:	Colon	Neutral	150 ms	Aside	clause (−0.20 st)
- — – --	Dash (hesitation)	— (rest on prior)	340 ms quiet	Hesitation	—
…	Ellipsis (hesitation)	— (rest on prior)	500 ms quiet	Hesitation	—

Verified with --explain on "wait... no, stop - now": ... appears as threecontrol:period · falling glide · pause 240 ms rows, while - appears as one hesitation-dash · quiet rest 340 ms row. In rendered audio, those three period rows collapse through the existing "first marker shapes, longest single pause wins" rule rather than creating three additive pauses; the remaining problem is that the boundary is still period-shaped, not ellipsis-shaped.

2.3 The collisions

• . vs !: same local glide + pause. ! can differ through utterance arousal and final TerminalFlourish, but it has no distinct exclamation boundary shape. Question already reads (it's the only riser, and it's the only mark that keeps pitch up).
• ... ⇒ period controls: the common ASCII ellipsis is explain-visible as three period rows and audio-visible as a period-shaped boundary, not one trailing-off gesture.
• - vs …: same quiet-rest mechanism, only the gap length differs.
• . mid-utterance: no role of its own; clause marks share one bucket — closure and continuation are not contrasted.

3. Recommendations for sound planning (VOICE_V9)

The organising idea: give each mark a distinct point in a (final pitch movement) × (intensity / finality) × (tail shape) space. Most of this can reuse existing planner and synth controls (role curves, final lowering, duration, pause overrides, brightness pressure, whistle/tension), but the work is not purely constant tuning: V9 needs a small per-marker boundary-signature model for glide magnitude/shape, plus a tail-shape directive for dash/ellipsis. Each recommendation below is independently implementable and testable inside a V9 branch; if any slice ships separately, it needs its own voice version and golden regeneration.

R1 — Split period from exclamation: settle vs punch

Today both fall. Differentiate by register, magnitude, intensity, and timing, mirroring the L-L% (statement) vs raised-peak H* L-L% (exclamation) split:

• Period = definitive settle. Deepen the final fall to the declination floor, increase pre_boundary_lengthening (slow the last syllable), and lower brightness/intensity on the terminal syllable. Give it a dedicated quiet, closing role (or a "settle" curve set) rather than falling through to ChattyReply. Suppress the whistle/flourish — a period is a statement, not a flourish.
• Exclamation = emphatic punch. Raise the terminal pitch peak and widen the range before the fall, steepen pitch_velocity, add a brightness/sparkle burst (brightness_pressure ↑, the event-sparkle path in synth.rs), and keep TerminalFlourish but flavour it as a punch (fall from a raised peak) rather than the question's rise. Avoid promising literal envelope-attack shortening unless V9 adds a bounded attack-shape control; today the attack envelope/transient durations are fixed voice constants. The existing archetype_tension lever already drives the flourish whistle — bias exclamation toward a high-energy fall, question toward the rise.

Acceptance feel: "now." vs "now!" vs "now?" must be three obviously different endings.

R2 — Normalise ... (and friends) into one ellipsis

Before tokenisation, or in a deterministic token-normalisation pass immediately after it, collapse runs of ASCII periods (..., . . ., ....) and the spaced/Unicode variants into a single Ellipsis HesitationMarker, and collapse --/---/spaced hyphens into one Dash. -- and the Unicode dashes are already handled (sequence.rs:126); the gap is specifically ... → ellipsis and multi-dot runs. This is the single highest-leverage change: it makes the ellipsis gesture actually reachable for the way people type, and it stops the period-control stutter. (Decide the boundary deliberately: !? / ?! interrobang runs should pick one terminal contour, not stack two.)

First reconcile the existing source-of-truth mismatch: either narrow FR-82/T-84 to the current Unicode-only implementation, or make ASCII ... part of the V9 audibility contract and update the historical V7 text to say it was planned but not implemented.

R3 — Make dash and ellipsis shape the tail differently, not just the gap

Keep both as quiet rests, but contrast the preceding syllable's tail so the break type is audible before the silence:

• Dash = abrupt cutoff. Truncate/clip the prior syllable's tail (shorten its release, hard amplitude gate), freeze the contour mid-move (a brief held suspension rather than a resolved glide), then a dash rest. Percept: interruption.
• Ellipsis = trailing off. Lengthen the prior syllable and apply a gentle amplitude decay, hold a level or shallow-falling (half-cadent) pitch — not a resolved fall to the floor — round off brightness, then a longer ellipsis rest. Percept: unfinished thought.

This recruits the SyllableTiming path the hesitation markers already own (with_pause_override, suppress_bridge, mark_hesitation, sequence.rs:142); the new pieces are a tail-shape directive (clip vs decay) and a held-vs-level contour on the hesitating syllable. V9 also needs to decide pause precedence: current V8 deployment can replace the marker's 340/500 ms pause with the role-gated ~930 ms hesitation turn gap before the next phrase, so the marker-specific rest length is not reliably audible in common "wait - no" / "wait … no" shapes unless the marker pause wins or the long-pause planner becomes marker-aware. The 340/500 ms gap difference should stay only as a secondary cue.

R4 — Contrast closure against continuation

Give the explicit period a real terminal identity and let clause marks read as "more coming":

• Period: low fall to floor + final lengthening + the R1 settle role.
• Clause , ; :: adopt a shallow continuation rise or level (L-H%) instead of the flat neutral glide, with the existing shorter pause. This makes a period audibly closed against a comma's open, the same way the question/period rise/fall pair already contrasts. This requires a new per-marker glide shape or a Continuation final-glide variant; today's SyllableFinalGlide only has Neutral, Rising, and Falling, with one fixed 3-semitone magnitude. (Optionally split : toward "presenting/listing" later — out of scope here.)

R5 — Strengthen the question rise (after glide-shape support)

The question already reads, but widen the terminal rise and add a small pre-final dip (L* before H-H%) so the rise is unmistakable on short final words, and confirm declination stays suppressed through the whole final segment (not just the last syllable). The confirmation is cheap; the wider rise and pre-final dip depend on the same per-marker glide-shape support as R4 because today's question glide is only the fixed 3-semitone Rising variant.

Proposed target signature (after R1–R5)

Mark	Final movement	Intensity / finality	Tail shape	Gap
?	strong rise, declination off	sustained	resolved up	long
!	fall from a raised, widened peak	loud / sparkle punch	sharp	long
.	fall to the floor	quiet settle, lengthened	resolved down	long
, ; :	shallow rise / level	sustained, light	resolved level	medium
—	contour clipped mid-move	held / abrupt	truncated	dash rest; long-gap gated
…	level / shallow hold	fades, lengthened	decayed	ellipsis rest; long-gap gated

4. Contract & sequencing notes

• All of R1–R5 are sample-affecting. If they ship as one batch, they require one VOICE_V9 bump, regenerated golden WAV hashes (DOOTDOOT_REGEN_GOLDEN=1 …), and a docs/validation/voice-v9-*.md acceptance note. If any recommendation ships before or after that batch, it needs its own voice version and golden fixture set. Do not ship any of them under V8.
• Order of work, lowest-risk first: R2 (normalisation — pure front-end, big perceptual win) → R4 (closure/continuation glides) → R1 (period/exclamation split) → R3 (dash/ellipsis tail shapes — needs a new tail directive) → R5 (question polish). R2 can land as its own task because it changes which markers fire, which every later recommendation depends on.
• TDD hooks: --explain already prints the per-token glide/pause/role and the curve grid, so each mark's signature is snapshot-testable at the planner level before any golden-WAV assertion. Add performance_planner / phrase_smoothing tests asserting the distinct glide + role + tail per mark, then pin bytes with goldens.
• Spec/plan sync: first fix the existing ASCII-ellipsis contradiction across FR-82, T-84, design.md §8.8, design.md §3.3, and code. Then add FR-* entries for the five-mark audibility contract and a Phase task block (T-NN) per CLAUDE.md's source-of-truth requirement.
• Acoustic validation: use scripts/acoustics to confirm "now." / "now!" / "now?" separate on dominant-peak range, 2–5 kHz share (exclamation punch), and final-pitch direction, and that "wait — go" vs "wait … go" separate on tail-energy decay and pre-gap duration.

Sources

• ToBI (Tones and Break Indices)
• Boundary tone
• Prosody (linguistics)
• Ellipsis
• Exclamation mark
• Speech Synthesis Markup Language (SSML)
• How is prosody controlled in modern TTS systems? (Milvus)