Reference Audio vs Description
Reference audio fixes timbre; words fix structure. When they conflict, protect structure first.
Related
A reference is not a master key
Give a model a reference and it brings the reverb, playing habits and era along with the tone, while ignoring your section plan. Reference and text split the job: the first owns timbre and surface, the second owns structure and events. Treat them as one control and they undermine each other.
Who fixes what
| Dimension | Reference audio | Text | Practice |
|---|---|---|---|
| Timbre | Strong | Weak to moderate | Fix tone with the reference |
| Space and reverb | Strong, often carried over | Moderate | Ask for dry if you need dry |
| Groove | Moderate | Moderate | Supply both and accept approximation |
| Instrumentation | Moderate | Strong | List the ensemble in words |
| Structure | Weak | Strong | Use section tags and duration |
| Key and harmony | Weak to moderate | Strong | Put exact requirements in text |
| Lyrics and diction | Weak | Strong | Text only |
Who wins in a conflict
| Conflict | Protect first | Move |
|---|---|---|
| Section count and duration | Text | Trim the reference or generate in parts |
| Lead timbre | Reference | Describe the ensemble, not a pile of tone adjectives |
| Tempo and key | Text, if contractual | Align in the DAW after generation |
| Reverb and space | Reference | Ask for dry separately or remove reverb later |
| Lyrics and diction | Text | A reference cannot control language |
Structure is the delivery skeleton, so sections, duration and lyrics come first; timbre is surface and can be replaced or regenerated. Tempo and key sit between the two, depending on whether they are hard requirements.
Combining the two inputs
Reference alone suits texture hunting and temp tracks; structure is cut later. Text alone suits a clearly specified form, with timbre left to adjectives and luck. With both, let text carry skeleton and exclusions, let the reference carry tone only, and state "reference for timbre only" in the prompt. That sentence stops the model from copying the reference's sections too. A single eight-to-fifteen-second excerpt usually obeys better than a whole track.
What neither can fix
Micro-timing, a real room and a singer's improvised delivery escape both controls. Hire a player or accept the generation and repair it. A reference is an anchor for style, not an acceptance test; acceptance returns to the section sheet, duration and lyric check.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Along the same line
Sources
- Copet et al., MusicGen（Meta AudioCraft, 2023）与 Stability AI 的音频条件生成技术说明。
- 录音制作教材中关于音色、空间与演奏细节的描述方法。
- 影视配乐沟通惯例，关于参考曲目在 Brief 中的角色。