AI music and audio failure atlas: ACE-Step Studio 3.8.4–3.8.5, caption timing and Piper streaming

Founder-affiliated synthesis, inspected 2026-11-07. Start with the symptom and exact engine/version; similar symptoms do not establish the same cause. Eight cases below paraphrase publisher release notes, one is our deterministic code fixture, and one is a documented integration limitation. We did not install Studio, generate music, reproduce the publisher fixes, measure VRAM or validate finished-media quality. No universal upgrade recommendation is implied. For a real check, preserve the existing input/output and environment, name the affected component and its version, and compare one permitted minimal case before and after the relevant change. Record failure as well as success. Download assets are not hardware validation: the inspected 3.8.5 release lists Windows x64, macOS arm64 and Linux packages; their existence does not prove your runtime works. When contributing a result, retain exact version, source revision, conditions, observed outcome and remaining uncertainty. A successful different backend or unrelated task cannot clear these cases. Current releases may supersede this dated map; use the linked primary record before changing a production workflow. ### 1. Lyrics become unintelligible later in a song Condition: Studio 3.8.0 regression; Blackwell CUDA capability 10.0+. Finding: Upstream attributes this to incorrect flash attention and reports automatic disabling on affected cards in 3.8.4. Evidence: upstream_report_not_reproduced. Source: https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.4 ### 2. Completed song disappears during final export Condition: Temporary directory unavailable or problematic path encoding. Finding: 3.8.4 reports in-memory FLAC encoding instead of a temporary file. This is a final-save path, not a synthesis-quality repair. Evidence: upstream_report_not_reproduced. Source: https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.4 ### 3. No karaoke timings when the vocal is buried in the mix Condition: Mel-Band RoFormer installed; original song mix. Finding: 3.8.4 reports extracting the vocal for alignment when the mix fails, including cloud recognition. No alignment accuracy measured here. Evidence: upstream_report_not_reproduced. Source: https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.4 ### 4. Saved maximum duration changes when loading a request Condition: Reuse, examples or writing assistant. Finding: 3.8.4 reports preserving the saved maximum duration. This does not guarantee that generated content satisfies the requested duration. Evidence: upstream_report_not_reproduced. Source: https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.4 ### 5. Automatic generated lyrics unexpectedly choose Russian Condition: Request leaves lyric language unresolved. Finding: 3.8.4 reports falling back to the application language. Explicitly requested language remains a separate condition to test. Evidence: upstream_report_not_reproduced. Source: https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.4 ### 6. Component update displays an unrelated explanation Condition: Audio.cpp CUDA libraries labelled with karaoke explanation. Finding: 3.8.4 reports separate component explanations. This is an update-screen usability correction, not an audio-output improvement. Evidence: upstream_report_not_reproduced. Source: https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.4 ### 7. Model or tool setup stalls when Hugging Face is unreachable Condition: Network cannot reach huggingface.co. Finding: 3.8.5 reports automatic fallback to hf-mirror.com. Confirm the actual download origin fits the operator's network and model-source policy. Evidence: upstream_report_not_reproduced. Source: https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.5 ### 8. LoRA training reports backend does not support Condition: RTX 20-series Turing; compute capability 7.5; ConvRot8 step. Finding: 3.8.5 reports an updated HOT-Step trainer with an F32 fallback for this step. It is not proof of training speed or memory fit. Evidence: upstream_report_not_reproduced. Source: https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.5 ### 9. A positive word interval collapses to zero after ASS timing conversion Condition: Reviewed words_to_ass.py revision 5496e6c05ff2135f6c60a8f7cb22637d34d092d2; 101–104 ms example. Finding: Our deterministic check observed both endpoints becoming 100 ms and the converter rejecting the event. Review quantized duration; do not label pre-conversion validation sufficient. Evidence: internal_deterministic_fixture. Source: https://veyrul.com/replies/rpl_01M6TVE4ZYHWK5D4FVKFS9ES21 ### 10. Sentence streaming is mistaken for continuous playback or a truncation fix Condition: OHF Piper Python API at 10880972db9e0aca101fed350eadf8f49deefe9f; consuming-player lifecycle. Finding: Saved source inspection shows sentence-level chunks and voice-specific PCM metadata. Test synthesis, queue drain and interruption separately; no Piper/Hermes reproduction or fix is established. Evidence: documentary_integration_limit. Source: https://veyrul.com/replies/rpl_01M6S9EFV5919ADJSB85VQSY5J

{
  "citations": [
    {
      "url": "https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.4",
      "note": "Publisher-reported fixes in 3.8.4, released 2026-11-05. Not reproduced by Veyrul."
    },
    {
      "url": "https://github.com/timoncool/ACE-Step-Studio/releases/tag/v3.8.5",
      "note": "Publisher-reported fixes and asset listing in 3.8.5, released 2026-11-06. Not reproduced by Veyrul."
    },
    {
      "url": "https://veyrul.com/replies/rpl_01M6TVE4ZYHWK5D4FVKFS9ES21",
      "note": "Founder deterministic code check; not an audio/render evaluation."
    },
    {
      "url": "https://veyrul.com/replies/rpl_01M6S9EFV5919ADJSB85VQSY5J",
      "note": "Founder source inspection; not execution."
    }
  ]
}