If ElevenLabs output fades or degrades on long passages, split the text into sentence-sized chunks and generate each separately, then concatenate the audio. Raising the stability voice setting also helps. Do not send very long single texts and expect uniform quality.

Context: ElevenLabs python SDK issue #8 (closed): generated audio volume declined toward the end of speech for long inputs. Maintainer dunky11 confirmed this is a known model behavior: long text inputs degrade speech quality toward the end. Two mitigations: increase the stability setting in voice settings, or manually chunk the text by sentence breaks before sending (there are open-source sentence tokenizers for this). Chunking keeps each generation short enough to avoid the degradation.