tts-voiceover-quality-gate

name	tts-voiceover-quality-gate
description	Use when generating, reviewing, or approving cloned or personal-voice TTS for paper explainer episodes before full audio, captions, or video render.

TTS Voiceover Quality Gate

Core Rule

Treat production voiceover TTS as a gated runtime workflow, not a one-shot render. The canonical chain is:

sample -> asr_diff -> human_approval -> full_tts -> merge -> captions -> render

Do not proceed to full_tts, audio merge, captions, or HyperFrames render until representative samples are approved.

Required Gates

• Prefer IndexTTS2 as the mainline when content consistency, terminology stability, clarity, and audio-video timing matter more than personal timbre similarity.
• Use IndexTTS 1.5 or CosyVoice as fallback engines; keep personal timbre cloning as an experimental branch until it passes the same content-consistency gates.
• Use a neutral 8-10s reference audio and reference text for IndexTTS2, IndexTTS 1.5, CosyVoice, F5-TTS, GPT-SoVITS, or similar engines.
• Block topic-specific reference text that can cause reference-text leakage, including phrases like 为什么重要 or reviewed script sentences.
• Generate representative samples first: seg_001, one mid-script technical segment such as seg_010, and the ending or CTA segment such as seg_014.
• Run ASR transcript diff when transcripts exist. If local ASR is missing, report that clearly and keep human listening review mandatory.
• Keep source_text and spoken_text separate. source_text is the reviewed narration; spoken_text may normalize numbers, English terms, formulas, and pronunciation.
• If delivery tuning is needed, run voiceover-emotion-coach after spoken_text is locked and before sample generation. Its default is preserve_original_ai_voice with low-intensity prosody; it must not change spoken_text or replace the original AI voice character.
• Apply the Pronunciation Normalization Contract before sample generation. High-risk phrases such as 更准确地说, adverbial 地, ambiguous 重, 按行归一化, QK^T, sqrt(d_k), and d_k must be rewritten or annotated in spoken_text so TTS does not guess the reading.
• Keep English technical and product terms as stable whole terms in spoken_text; for example ChatGPT, Claude, token, Attention, softmax, KV Cache, and Multi-Head Attention must not be split, translated, or phonetically rewritten unless the source script explicitly requires spelling an acronym.
• EP05-derived high-risk terms must be checked before sample or full generation when present: naked Q, K, V, Q/K, KV cache, mθ_i, nθ_i, δ_i, θ_i, n-m, and polyphonic phrases such as 长上下文.
• Run duplicate and near-duplicate checks before TTS so repeated narration is blocked before audio exists.
• Import only approved, postprocessed WAV files into the canonical audio/voiceover.wav slot.
• For segmented TTS, require an audio freshness check before merge or final render whenever the manifest, spoken_text, or critical segment list changes.

Stop Conditions

Stop and request human review when:

• Sample audio repeats words or leaks reference text.
• English terms such as ChatGPT, Claude, token, Attention, KV Cache, or vLLM are skipped, split incorrectly, phonetically rewritten, drift between segments, or unclear.
• Formula phrases such as Q 乘 K 转置, 根号下 d k, softmax, or Multi-Head Attention are misread, swallowed, or split into unintended fragments.
• Audio contains electrical noise, clipping, long silence, or obvious segment discontinuity.
• ASR transcript diff shows drift from spoken_text or detects leak-prone phrases.
• review/sample_audio_review.json.status is not approved for the current segment text hash.
• A critical segment wav is older than the current manifest, missing, invalid, or produced by a failed batch rerun.

Verification

Use the existing project commands when available:

npm run voiceover:duplicate-guard
npm run audio:f5-check-reference
npm run audio:indextts2-prepare-segments:ep02
npm run audio:indextts2-generate-samples:ep02
npm run audio:f5-generate-samples
npm run audio:asr-transcript-diff
npm run audio:sample-review-gate
npm run audio:pronunciation-gate:ep05
npm run audio:freshness-gate:ep05-critical

Only after those gates pass should the workflow continue to full_tts -> merge -> captions -> render.