How this demo works.
SongLLM is a concept: a layer that listens to a song, points at the seconds that fail, says why in a producer's words, and hands the fix to the tool you already use. This site shows that concept working on three tracks, end to end, with today's tools. It is honest about where the edges are.
What is real
| Piece | How it was made |
|---|---|
| The songs | Generated with Suno from written lyrics and style prompts. Three picked from a batch of eight for contrast: a four-minute instrumental, a rap cypher, and a whispered sketch. |
| The waveform | Peak amplitude of the actual audio, 720 bins, drawn as SVG. The dashed lines are section boundaries found by novelty analysis on chroma, timbre, and loudness, then labeled by hand against the model's reading. |
| Track details | An audio-native model (Gemini 3 Flash) heard each file and answered four questions: full technical reading, where it sags, what the hook is, and three changes. Its answers were cut for length and corrected where the file proved it wrong: a mis-heard lyric quote, an outro timestamp past the end of the track. The unedited text is on the main page under "Track details", in the "Unedited model output" fold. |
| Tempo and key | Tempo is measured: a comb filter over the onset envelope in 0.005 BPM steps, cross-checked by autocorrelation, with the bar line placed by hand where the kick enters. It is labelled "measured" on the page. Key is the model's call, labelled "model", or "disputed" where the chroma profile disagrees; the disagreement is shown, not resolved. |
| Loudness | Integrated loudness, loudness range and true peak from ffmpeg's EBU R128 meter on the same file you hear. Labelled "measured". Nothing in the reading's prose about loudness is a measurement unless it carries that label. |
| The bar ruler | The small ticks under the waveform are every four bars of the measured grid. The artwork videos were cut on the same grid, so every cut lands on a bar line and the loops restart in phase. |
| The notes | Edited from the model's answers, pinned to the timecodes it cited. Click one and the audio jumps there. |
| The changes | The model's three changes, each as a card: what was heard, what moves, what stays, then one instruction for the destination you pick (Suno Studio's chat bar, an ElevenLabs Music composition-plan instruction, or DAW moves). Copy and open are separate actions; nothing is sent anywhere. |
| The one that is applied | Control, change 1. The last beat before each chorus (bars 33, 67, 94 on the measured grid) is muted with a 5 ms fade and a tuned sub impact (98 Hz gliding to 49 Hz, G1) lands on each downbeat. Done as a numpy edit on the decoded bounce, encoded with ffmpeg. No EQ, no normalisation, no limiting. The A/B plays the same eight bars from both files on one clock; the revision measures 0.3 LU quieter over the excerpt because one beat is silent. The other eight changes are instructions only. SongLLM does not edit songs; this one was applied by hand so the advice can be heard. In a live build the before and after comes from you: make the changes in your own tool, upload the new version, and SongLLM compares it with the last notes (coming soon, not built). |
| Checked against the signal | Every timed claim in the reading that says something enters or leaves ("the kick arrives at 0:31", "the bridge removes kick and snare") is measured on the file. The audio is split into its percussive and harmonic parts; drum claims read the percussive part, melodic claims the harmonic. The named frequency band is compared four seconds before and after the moment. A claim holds if it moves the stated way by 3 dB or more. If the move only appears a few seconds away, the page says where. If it does not appear within eight seconds, the claim is marked "not found in the signal" and left on the page. The first pass across the three tracks found 6 of 19 claims holding, 6 a few seconds off, and 7 not in the audio. Two of the misses sat under suggested fixes (the Control bridge and the Big Lab Flex drop), so those notes and fixes were rewritten from the measurements; the corrections ledger shows what changed. Now 8 of 18 hold, 5 are a few seconds off, and 5 are not found and stay on the page, marked. That is the point of the check: the model's timestamps are approximate and some of its structural claims are wrong, and a reader should know which. The spectrogram under the waveform is the same analysis, drawn. |
| Key and pitch | Key stays disputed on all three tracks because audio-language models do not yet hear pitch reliably. The PitchBench study (arXiv 2605.26176, 2026) tested current models across 28 conditions and found unstable pitch perception even on clean synthetic notes. A live build would read key and chords from signal tools and use the model for what it does well: arrangement, energy, mix and meaning. |
| The words | The same model heard each file with its lyric sheet and was asked, as a songwriter, what the lyric seems to say, how it measures against named songwriting conventions, other ways to say a line, title options, and every line a careful listener might want to look at twice (explicit language, double meanings, cultural references, real people, echoes of other works). It ran twice; flags from both runs are merged and each says how many runs raised it. One flag on Big Lab Flex was added by the editor because both runs missed it, and it says so. Syllables are counted by rule, not by the model. Nothing here is the final word: suggestions are options, and flags are questions for the writer. |
| Made with | Pick the tool the song came from and every fix rewrites itself for it: AI music tools (Suno, Udio, ElevenLabs Music, Google Gemini with Lyria, Stable Audio, Mureka, or another), recording software (Ableton Live, FL Studio, Logic Pro, GarageBand, Pro Tools, Cubase, Studio One, Reaper, Bitwig Studio, BandLab), or plain steps for whoever mixes it. Each fix exists in three forms: a prompt for AI tools, section-editing language for ElevenLabs, and steps for a DAW. The note under each fix only says what that tool is known to do. |
| Release check | Each song is held to published standards first (Spotify's loudness and true-peak guidance, Spotify Canvas specs, the distributor rule for explicit tags), then industry practice (a stream counts after 30 seconds, so where the hook lands matters), then craft, then trend and taste: what may already feel played out, from lyric cliches to overused sounds, labelled as the artist's call. Anything outside that list is creative choice. Sources are linked on each row. |
| The video | The same kind of model watched each demo video with the song's audio and was asked to judge it like a label's video commissioner judges a first cut: what works, what to improve, each with a time, and the video it would commission for the song. The demo videos were made quickly, not by a crew, and the prompt says not to go easier on them for that. Every cited time was checked against the frames before it shipped; one mistake was corrected (a verse named after the wrong lab). Video uploads are coming soon. |
| Lyrics prompt | The lyrics section builds a prompt you can paste into ChatGPT, Claude, Gemini or any assistant: your lyrics, what the song seems to say, the lines worth a second look, and the suggestions. It asks the assistant to help you decide, keep your voice, and match syllable counts, not to rewrite the song. |
| Check your song | The upload form is the full product's front door. In this demo nothing leaves your device: the page checks the file name, type (MP3, WAV, AIFF, FLAC, M4A) and size (up to 50 MB), never opens the file, and clears the form when you close it. You confirm you made the song or hold the rights. Then the page simulates the check: it walks through the stages a live build runs (listening, measuring, reading the lyrics, checking notes against the audio, writing fixes, the release check) and opens the finished report on the closest sample song, marked as a simulation. Nothing is generated while you wait; the reports were prepared once. A live build would add: encrypted upload to private storage behind short-lived signed links, a malware and format scan before processing, deletion after the check, no training on your audio, no public links, and a rate limit per account. |
| Corrections | Every place a human or a tool overrode the model is listed under "Corrections and provenance" on the main page: the tempo values, a mis-heard lyric quote, a timestamp past the end of the file, and the key dispute. |
What is scoped
Upload does not upload. The questions are prepared. Free text is routed by intent and the router fails closed: a question about structure, weak spots, the hook, the changes or the comparison gets the prepared answer; a question about loudness, tempo, key or length gets the measured numbers; a question about the mix, the instruments, the mood or the genre gets that paragraph of the reading; a question naming a tool points at the destination switch; a question this demo cannot answer (stems, stereo width, frequencies, generating or editing on request, transcription) says so and gives the reason; anything else is marked "not sure" and offered the five questions. It never falls back to the overview. The readings were produced once and stored. Nothing on this page calls a model while you use it.
What a live build does
The same questions run on your file when you upload it, with a music-understanding model behind them and the measurements taken by tools alongside. Open weights exist for exactly this task (NVIDIA's Music Flamingo, released with the UMG collaboration in January 2026), and the audio-native frontier models can already do it, as this page shows. The reading pins to the waveform with every number labelled measured, estimated or model; the changes render as cards with what they protect; the handoff opens the tool; and when you upload your revised version, SongLLM compares it with the last notes: what improved, what got worse, what is still open. Everything on this page is the interface for that.
Why the critic and not the editor
Suno Studio already edits by chat. Udio and ElevenLabs will follow. The generators own the refine loop. What none of them ship is an independent reading that points at the second and says why, in words a producer would use, and that works across every tool. Universal Music Group and NVIDIA framed their January 2026 collaboration around exactly that: understanding, discovery, and attribution. This demo is a thin, honest version of the same idea, from the outside.
Rights and disclosure
The songs are Suno outputs and are presented as such. The domain, the concept, and this demo are available to the right team. Talk to us.