FAQ

Does the fast test suite prove real music generation works?

No. The current unit tests mostly validate abstraction and tensor plumbing with fakes. Real model smoke tests are planned and must be opt-in because checkpoints are large.

Which model should be the default?

The base default is acemusic, a lightweight remote ACE Music API backend, because the default package profile should not install local ML runtimes. For local generation, ACE-Step Diffusers XL Turbo through acestep is the supported Apple/GPU path that passed the current 30-second reference-floor smoke.

elevenlabs is the second supported remote backend, but it is explicit rather than the default. Two remote endpoints are enough for the lightweight package profile; new provider work should now focus on local/open-weight engines unless a remote endpoint unlocks a missing abstraction.

Does AbstractMusic use Suno, OpenAI, or Anthropic by default?

No. OpenAI exposes speech/audio APIs but not a text-to-music endpoint suited to this backend. Anthropic does not expose music generation. Suno has a consumer music product, but the public API surface found during review is third-party or unofficial, so it is not a default backend. A Suno adapter can be added later if there is an official, stable, documented API contract.

Does the ElevenLabs backend include voice?

No. AbstractMusic calls only ElevenLabs Music endpoints. Voice, text-to-speech, speech-to-speech, and voice cloning belong in AbstractVoice.

How do I specify instruments or vocalist traits?

Use plain text in the prompt (for example: “jazz with saxophone and brushed drums”) or set provider-neutral style tags:

• CLI: --style "saxophone, brushed drums, female vocalist"
• REPL: /style saxophone, brushed drums, female vocalist

This nudges the model, but it is not a guaranteed identity selector (no voice cloning here).

Can HeartMuLa generate music?

Yes. The HeartMuLa model card describes it as a text-to-audio/music model, and the official HeartMuLa repo documents lyrics and tag conditioned generation with HeartCodec. It is Apache-2.0 licensed, but it is a heavy optional provider candidate rather than a base dependency.

Why not make Omni2Sound a default provider?

Omni2Sound is CC BY-NC 4.0, very large, CUDA/script-oriented, and focused on video/text-to-audio and foley-style generation. It should not be selected silently for commercial-capable music generation.

How does AbstractMusic decide which providers are available?

By checking what can run right now, without loading a model. A local provider is available when its runtime extra is installed and at least one of its models already has weights in the Hugging Face cache. A remote provider is available when an API key is configured and its endpoint answers a probe; all remote providers are probed in parallel under a 5-second deadline.

available_providers(...) therefore lists only what you can generate with immediately. To see the full picture, including providers you could enable and models you could download, use provider_details(...), which reports every provider with a usable flag and the reason it is not. The packaged registry remains the complete catalog: abstractmusic repl then /models lists every known model regardless of what is installed.

When does AbstractMusic rewrite my prompt?

Only when you opt in. Your prompt passes through unchanged unless you request enhancement (--enhance-prompt) or enable a feature that needs a caption: long-form structure maps (durations of 45 seconds or more, on by default) or --auto-lyrics. For checkpoints the registry marks caption_sensitive (the ACE-Step XL sft/base variants), feature captions stay compact — your prompt plus the section map, no template prose — because long template captions degraded those checkpoints in validation. The planner records what it did in provenance warnings, printed to stderr and stored in generation metadata; --print-plan shows the exact text sent to the model.

Do we use 8-bit models?

When official 8-bit model artifacts exist, they should be preferred. If no official 8-bit artifact exists, use the smallest official 16-bit path. The registry records whether a reviewed model has an official 8-bit source.