Common causes: inconsistent API params (voice ID/model, sample rate, quality preset), rate-limits or too many parallel requests, inconsistent SSML/prompt formatting, or lossy post-processing. Best practices: lock the same voice ID and sample rate, output lossless WAV, throttle or queue batches (small pauses), keep prompt templates identical, monitor API errors/retries, and regenerate any glitchy files. If it persists, collect request IDs and logs for support.

Likely causes: API throttling/concurrency, switching voice settings or voice_id mid-run, inconsistent sample_rate/format, using streaming endpoints, or short/context windows causing variability.

Practical fixes: lock a single voice_id + voice_settings for the entire batch; use the synchronous (non-stream) endpoint; limit parallel requests or add 100–500 ms spacing; keep sample rate/codec identical; add tiny silence padding and normalize loudness in post; test a 5–10 file subset first.

See ElevenLabs docs for batching and concurrency tips.