News · ai music — By WATMM Desk —
New guitar-synthesis paper lets you edit bends and dynamics string by string
An arXiv paper from Hsueh-Wei Fu, Yen-Tung Yeh and Yi-Hsuan Yang turns MIDI into guitar audio in two steps, first predicting editable pitch and loudness curves for each string, then rendering them to sound. The checkpoints are free to download, and listeners rated it above two comparison systems.
Key facts
- What: a two-stage guitar synthesiser posted to arXiv on 27 Oct 2026 and submitted to ICASSP 2027. It predicts per-string pitch (f0) and amplitude curves from MIDI, then renders them with latent flow matching.
- Free: a demo page with audio, code and model checkpoints is public; the Hugging Face checkpoints are tagged CC BY 4.0.
- Why it matters: MIDI says which notes a guitarist plays but not how. Editable curves are a route to programmed guitar parts you can actually shape, not just regenerate.
A new research system turns MIDI into guitar audio while letting you edit how each note is played, string by string. The paper, posted to arXiv on 27 October by Hsueh-Wei Fu, Yen-Tung Yeh and Yi-Hsuan Yang, is a research model, not a plugin, but the audio demos and model checkpoints are free to try.
What's free
The project page lets you listen to and A/B the outputs in a browser, and links to the code and to checkpoints on Hugging Face, which are tagged CC BY 4.0. Running it yourself means a Python research setup, not a DAW install. If you just want guitar in a track today, recording guitar at home is still the cheaper route, and our MIDI generators in your DAW guide covers tools you can use now.
What the researchers did
The problem they name is simple: per the abstract, timed MIDI events "decide what is played but not how". Pitch inflection (bends, vibrato) and per-string articulation are left unspecified, so a model that maps MIDI straight to audio has to invent the performance itself, and you get little say over it.
Their fix splits the job in two:
- A performance predictor takes the note events and note-level dynamics and predicts a pitch (f0) curve and a loudness (amplitude) curve for each of the six strings. These curves are editable.
- A renderer turns those curves into audio using latent flow matching, a generative technique that works in a compressed audio representation. The project page describes it as producing six-string audio through a neural audio codec.
According to the authors, the system beat a matched single-stage version and DDSP-Guitar, an earlier guitar synthesiser, on both control fidelity and perceived quality, while keeping the controls editable. The paper has been submitted to ICASSP 2027, so it hasn't been through that conference's peer review yet.
What we think
The interesting part isn't "AI makes guitar sounds". It's the middle layer. Programmed guitar usually fails in the bends, slides and dynamics between the notes, and most AI audio tools give you a regenerate button rather than a knob. Here the performance is exposed as curves you could, in principle, draw on like automation, per string. That is the shape a producer-facing tool would need.
The caveats are real. This is a guitar-only research model, judged in the authors' own listening tests, with no plugin, no DAW integration and no word on real-time use. We'd treat it as a signpost: the useful direction for AI instruments is editable intermediate controls, not one-shot audio. For the wider split between generating notes and generating sound, see text-to-MIDI vs text-to-audio.
Sources
- Expressive Guitar Synthesis via Controllable Latent Flow Matching, arXiv (primary source). Accessed 29 Oct 2026.
- Expressive Guitar Synthesis via Controllable Latent Flow Matching (demo page), Hsueh-Wei Fu et al. (primary source). Accessed 29 Oct 2026.
- ctrl-lfm-guitar-synth model checkpoints, Hugging Face (primary source). Accessed 29 Oct 2026.
Researched with AI assistance, written and checked by WATMM Desk.