Higgs Audio v3: condition on a tail-prepared reference to stop end-of-clip hiss

[niksedk]

Follow-up to #14546 and the "end noises" still heard after it (also the 60 % estimate in #14480 comment).

What the re-test found

Re-tested on the shipped 2026-09-05 audio.cpp build (Metal, Q8_0) with an instrumented local build of the same commit:

• The More refact in Utilities #454 codec fix works as designed. On identical seeded codes, no tail context vs 8 frames drops the last frame by 8-12 dB, and 64 frames of context gives the same samples (max diff 0.005). Nothing more to gain there.
• The remaining hiss is in the codes, not the codec. Holding the final frame's codes for 8 extra frames decodes to a steady hiss at the same level (-26 to -35 dBFS), so the language model itself emits noise codes in its last 5-7 frames before the end-of-audio token. The raw codes carry no stray BOC/EOC, the delayed tail is the normal EOC staircase, and sglang-omni's reference sampler has the identical delay/EOC state machine. Temperature/top-k/top-p changes do not help.
• What decides it is how the reference clip ends. The model continues the reference in context and ends its own clip the same way. Sophie_Anderson.wav ends at -19 dBFS mid-noise and every clip from it ended hissy; female_06.wav ends at -83 dBFS and its clips ended clean.

Fix

CloneReferenceTail.PrepareAsync makes a cached copy of the reference (voices/prepared/<name>.wav, keyed on size + mtime + recipe version): trailing silence/noise trimmed at the pipeline's peak-relative threshold, 50 ms fade-out, 400 ms of digital silence appended, 24 kHz mono. Higgs Speak conditions on that copy; the transcript is still read from the original's .txt. Falls back to the original when ffmpeg is missing or fails.

Seeded A/B, 48 clips per treatment (4 voices × 12 seeds, same seeds per column), level of the loudest of the last four codec frames:

reference	median	loud endings (> -40 dBFS)
as is	-46.3 dBFS	18 / 48
trim only	-55.8 dBFS	5 / 48
trim + fade + pad (this PR)	-64.3 dBFS	3 / 48

No runaway generations in 192 prepared-reference requests. The remaining loud endings are Morgan Freeman's clip, whose own room tone the model copies. Pad without fade was rejected: an abrupt cut followed by silence made the model emit one loud burst before its silence.

Also retries once on audio.cpp's Higgs TTS generation reached max_tokens before EOC (one in ~370 requests observed; no seed is sent so the retry samples afresh), which previously failed the segment outright.

Tests

• Parameter builder: filter order (trim and fade on the reversed signal, pad after), invariant formatting, clamping.
• CloneReferenceTail: prepared path, stamp, cache reuse, missing-reference fallback, runaway-error matcher.
• ffmpeg-backed: real preparation of a synthetic tone + noise-tail clip: noise tail gone, 400 ms of zero samples appended, fade before the pad, second call reuses the copy.

🤖 Generated with Claude Code