Raw-speech truncation bound — the oracle feature extractor's default
(audio_feature_extractor.py call: truncation=True,
max_length=480_000 → w[:max_length], i.e. 30 s at 16 kHz) applied
BEFORE feature extraction. Mirrored in decodeAudio so a >30 s clip
yields exactly the capped 750 soft tokens AND the matching mel frame
count — the prompt builder hard-checks that the tower's output length
equals the splice's soft-token count.
Raw-speech truncation bound — the oracle feature extractor's default (audio_feature_extractor.py call: truncation=True, max_length=480_000 →
w[:max_length], i.e. 30 s at 16 kHz) applied BEFORE feature extraction. Mirrored in decodeAudio so a >30 s clip yields exactly the capped 750 soft tokens AND the matching mel frame count — the prompt builder hard-checks that the tower's output length equals the splice's soft-token count.