The spoken text.
Without the long tail.

Hear the same take before and after trimming

Sixteen diagnostic takes span both languages, both models, both timing arms and both generation methods. The 150-ms crop is the standard preview; 100 and 200 ms test the ending. A long raw tail is not hidden. All players are exclusive.

Reassembled conversations

The same selected candidates as in the original experiment, now with script-anchored cuts and gentle fades. Compare the previous 400-ms montage with the new one. Missing endpoint verification is labelled; it is not filled with an unsafe cut.

Exactly what was edited

This is waveform post-processing of the same complete generation pool. No model, prompt, seed, identity reference or selected candidate was changed. All 1,536 original raw takes and the original measured experiment remain available.

1. Reuse the cached Parakeet Redux word timestamps from the completed GPU evaluation. No additional ASR, EARS inference or GPU job is needed.
2. Match the literal requested script to the recognized words. Use the first and last verified script word as boundaries, rather than the last arbitrary ASR word, which may belong to later mumbling.
3. Keep up to 80 ms before the first word. Keep 150 ms after the last word by default; selected diagnostic examples also compare 100 and 200 ms. Natural tail audio is retained where available, then faded. It is not assumed to be pure silence. If the source stops sooner, only zeros are added to fill the requested tail.
4. Apply a gentle sine-squared fade-in of up to 20 ms and cosine-squared fade-out of up to 50 ms, strictly outside the verified word span. Do not modify samples inside the spoken-word interval. Do not time-stretch, regenerate or alter internal block pauses.
5. Reassemble from lossless 48-kHz masters and encode the full conversation once. The gap is exactly 400 ms from the previous last word end to the next first word start, including retained tail/lead guards; added silence makes up only the remainder.

Conservative editing QA requires both script endpoints, at least two matching tokens, at least half the literal tokens matched and every intended turn aligned. If an endpoint is not recognized, that candidate is marked for review and remains available as raw audio. No guessed cut or silent fallback appears in a new montage. This added editing check does not alter the original experiment's generation/scoring gates. It can reject otherwise audible takes when ASR misses a word.

This speech-boundary pilot can remove intentional wordless sounds after the final spoken word; it does not promise full preservation of terminal laughter, gasps or sighs. Nonverbal events between the first and last scripted words remain. Keep the raw recording when such an ending matters.

Scores: these edited derivatives have not been rescored. Do not copy the raw audio's Content Enjoyment, Genuineness or WER onto them. Shorter tails alone do not establish better identities or artistic performance. The original experiment's separate measurements are linked for reference.

Both models retain the same genuine, warm, realistic, pleasant-to-listen-to prompts and the same two German reference voices for English and German. The default conversation view uses independent single-speaker turns; four/five-turn blocks preserve their internal timing. A pause before the next independently generated block does not repair a switch error inside that block.

Editing statistics · Every candidate's crop coordinates · Every episode's source IDs and join clock · Large 1.7 model