TEMPO · KEY · PITCH

Melody in your ears,
notes in your hands.

Drop in music to see how the beat moves and how high the melody goes.
Check its BPM, key and estimated vocal range to quickly see whether the song suits your voice.

01Your audio

Up to 20 min · 100 MB

Drop your music here

Audio files or MP4/MOV video soundtracks.

WAV · MP3 · FLAC · M4A · OGG · AIFF · AAC · MP4 · MOV

Audio stays off the cloud. Temporary uploads are deleted after analysis.

Analysis engine

(Recommended) Analyze on this device without uploading audio. Switching engines analyzes the selected range.

Enhanced vocal analysisOptional

Separate vocals with htdemucs, then estimate notes with GAME Large. Downloads about 568 MB of extra models on first use and uses more time and memory. Browser mode requires WebGPU. Changing this option reanalyzes the selected range; tempo and key still use the original mix.

Local Python setup & troubleshooting

Local Python runs on the same computer as your browser. Follow the macOS, Windows (PowerShell), or Linux instructions below. Set it up once; if already installed, jump to step 3.

How the browser communicates with Python

Uvicorn runs the local FastAPI HTTP service. The browser sends audio with POST /api/analyze (multipart/form-data), receives JSON results, then downloads MIDI with GET /api/download/{token}. Audio is sent only to the service on this computer.

1. Prepare your project folder

Download or clone the complete project, including web_app.py and requirements.txt. Open Terminal on macOS/Linux or PowerShell on Windows, then run cd "path/to/music-detection" with your actual folder path. Run all commands from that folder.

2. Install Python dependencies once

Choose your operating system below. Use Python 3.12 in the project's .venv; these commands preserve an existing environment. Wait for installation to finish. After installing tools, reopen your terminal and return to the project folder before continuing.

macOS

With Homebrew installed, install the required tools:

brew install uv ffmpeg git

Homebrew installation guide ↗

Reopen your terminal, return to the project folder, then run:

uv python install 3.12
test -d .venv || uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python -r requirements.txt

Windows 10/11 · PowerShell

Install uv, Git, and FFmpeg with WinGet. If winget is unavailable, install or update App Installer using the Microsoft guide below.

winget install --id astral-sh.uv -e
winget install --id Git.Git -e
winget install --id Gyan.FFmpeg -e

Microsoft / WinGet ↗

Reopen your terminal, return to the project folder, then run:

uv python install 3.12
if (-not (Test-Path .venv)) { uv venv --python 3.12 .venv }
uv pip install --python .\.venv\Scripts\python.exe -r requirements.txt

Linux · Ubuntu / Debian

These commands are for Ubuntu/Debian. On other distributions, install git, curl, and ffmpeg with your package manager, then install uv.

sudo apt update
sudo apt install -y git curl ffmpeg
curl -LsSf https://astral.sh/uv/install.sh | sh

Reopen your terminal, return to the project folder, then run:

uv python install 3.12
test -d .venv || uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python -r requirements.txt

uv / Python ↗

3. Start the local service

Run the command for your operating system. On macOS you can also double-click start_ui.command in Finder. You only need to start the service each time; no dependency reinstall or virtual environment activation is required.

macOS / Linux

.venv/bin/python -m uvicorn web_app:app --host 127.0.0.1 --port 8765

Windows · PowerShell

.\.venv\Scripts\python.exe -m uvicorn web_app:app --host 127.0.0.1 --port 8765

The service is ready when Terminal shows:

Uvicorn running on http://127.0.0.1:8765

Keep that Terminal window open while analyzing. Press Control + C in that window when you want to stop the service.

4. Return here and analyze

Select Local Python and add an audio file to start whole-track analysis automatically. To analyze a clip, adjust the range and press Analyze selected range. Switching engines analyzes the current range. After starting the service, press Analyze again to retry a failed attempt.

Beat/key weights download on first use. Standard mode uses the bundled GAME Small model. Enabling enhanced mode downloads htdemucs and GAME Large (about 568 MB extra). After updating, reinstall requirements.txt and restart FastAPI. Temporary audio is deleted after analysis; MIDI downloads last up to one hour.

Open the local service page ↗

That page also defaults to Browser ONNX. Select Local Python there if you want to use the Python service.

The service has been tested locally on macOS. These Windows and Linux setup instructions have not yet been verified end to end on those systems.

Still unable to connect?

Connection failed: check that the service is running on this computer at port 8765. Open the local service page above. If it does not open, check the Terminal error first. You can switch to Browser ONNX while resolving the issue.
ModuleNotFoundError or a missing .venv: complete step 2. Use .venv/bin/python on macOS/Linux or .\.venv\Scripts\python.exe on Windows. If requirements.txt is missing, return to the project folder in step 1.
FFmpeg not found: install it using the commands for your OS in step 2, reopen the terminal, and check ffmpeg -version. Address already in use means port 8765 is occupied; use the running service or stop its terminal session with Control + C before restarting.

Connecting from GitHub Pages or another website

Allow your site's exact origin when starting the local service. Replace YOUR-NAME with your GitHub account; do not include the repository path. If the service is already running, stop it first and restart with this setting.

macOS / Linux

TEMPO_ALLOWED_ORIGINS=https://YOUR-NAME.github.io .venv/bin/python -m uvicorn web_app:app --host 127.0.0.1 --port 8765

Windows · PowerShell

$env:TEMPO_ALLOWED_ORIGINS="https://YOUR-NAME.github.io"
.\.venv\Scripts\python.exe -m uvicorn web_app:app --host 127.0.0.1 --port 8765

Your browser may restrict access from a website to a local service. If blocked, use the local service page or Browser ONNX. Online GitHub Pages access has not been verified for this project.

02Analysis results

Awaiting audio

TEMPO

—BPM

Follows detected beat timing; a higher click marks each detected downbeat.

TIME SIGNATURE

Beats per bar

—

MUSICAL KEY

Global key estimate · S-KEY

—

Detected vocal range

GAME Small

Note estimates for this track or selection.

Estimated from this recording, not the singer’s full vocal range. Harmony and instruments may be included; confirm by listening.

Detect BPM automatically. For variable tempo, see the average and export a tempo map.

Tempo plus Lead Vocal notes when available

01 / DROP

Add your audio

Choose a file, or drag and drop.

02 / DETECT

Discover the rhythm

Find beats, tempo, meter and key.

03 / CREATE

Back to creating

Download a .mid file for your DAW.

BEHIND THE MUSIC

Meet the models

Specialized models for rhythm, key and singing pitch, with optional vocal separation before enhanced pitch analysis.

Beats & downbeats

Beat This!

CP JKU · ISMIR 2024

Detects beats and the first beat of each bar. This app uses their timing to calculate constant or average BPM, estimate meter, and generate a MIDI tempo map.

Half-time, double-time or missed beats can occur. A variable tempo map also reflects jitter in detected beats.

Source code Research paper

Global musical key

S-KEY

Deezer · ICASSP 2025

Uses self-supervised learning to estimate a global key from 12 major and 12 minor keys. This app analyzes audible clips of at least 3 seconds.

Estimates one global key, not chords or a modulation timeline. Relative keys and key changes can be ambiguous.

Source code Research paper

Singing pitch & note range

GAME Small / Large

openvpi · v1.0.3 · FP32

GAME Small estimates singing notes from the mix by default. Enhanced mode uses the official GAME Large v1.0.3 ONNX release on separated vocals. Both display the note timeline and lowest/highest accepted pitches.

Not a lead-vocal separator: harmony, accompaniment and octave errors can affect the range. Notes shorter than 80 ms are excluded. No labeled vocal-range accuracy benchmark has been completed.

Weights: CC BY-NC-SA 4.0 (attribution, noncommercial, ShareAlike). Code: MIT. Credit: openvpi/GAME contributors; original model release by yqzhishen.

Source code ↗Model weights ↗Official Large release ↗Model license ↗Original model & credits ↗

Vocal separation · enhanced mode

htdemucs

Meta · Demucs / Ghilda · ONNX

One htdemucs model separates vocals before GAME Large analyzes them. Separation memory is released before pitch analysis. The browser uses Ghilda’s community ONNX export of the official weights (about 174.3 MB / 166.2 MiB).

The vocal stem can contain backing vocals and leakage. Separation artifacts may change notes; enhanced mode is not guaranteed to improve accuracy.

Demucs and this ONNX export: MIT. GAME Large weights: CC BY-NC-SA 4.0.

Source code ↗Community ONNX export ↗

MIDI combines Tempo and available GAME notes in a Lead Vocal track; otherwise it contains tempo only. GAME does not isolate the lead singer. The key is shown in the results.

ONNX vs. PyTorch: model differences

Both engines analyze tempo, key and pitch. Browser mode uses ONNX; Local Python uses PyTorch for Beat This! and S-KEY, and ONNX Runtime for GAME and optional htdemucs.

Recognition accuracy

Model weights: FP32

Beat This! and S-KEY exports have PyTorch conversion checks. GAME Small and Large use official FP32 ONNX. Neither enhanced-mode gains nor vocal-range accuracy have been benchmarked here. Separation can retain backing vocals or introduce artifacts, so it does not guarantee improvement. FP32 is numerical precision, not recognition accuracy.

Model numerical precision

Model	Browser ONNX	Local Python
Beat This!	FP32ONNX Runtime Web	FP32PyTorch
S-KEY	FP32ONNX Runtime Web	FP32PyTorch
GAME Small / Large	FP32ONNX Runtime Web	FP32ONNX Runtime
htdemucs	FP32ONNX Runtime Web · WebGPU	FP32ONNX Runtime · CPU

Beat This!, S-KEY, GAME Small/Large and htdemucs use FP32 floating-point weights. These bundles are not FP16 or INT8 quantized. FP32 describes numerical precision, not a recognition accuracy percentage.

ONNX export error · maximum absolute difference

Beat This!≈ 6.1393 × 10⁻⁶

Beat/downbeat logits, tested at 63, 128, 1,264 and 1,500 spectrogram frames.

Validation record

S-KEY≈ 3.7551 × 10⁻⁶

24 key scores, tested on clips of 3, about 25 and about 75 seconds.

Validation record

These are the recorded maxima for the pinned Hugging Face exports, comparing PyTorch with ONNX Runtime on CPU using identical inputs. They are not BPM errors, accuracy percentages or guaranteed bounds for other audio, exports or browser/GPU runtimes.

Browser ONNX · Recommended

Getting started: Open this page; no Python installation needed. Models download on first use and are cached when browser storage is available.
How it runs: ONNX Runtime Web runs in a Web Worker. Standard Beat This! and GAME Small try WebGPU with WASM fallback; S-KEY uses WASM. Enhanced htdemucs and GAME Large require WebGPU and run in sequence.
Audio processing: Audio is decoded and analyzed in your browser. It is not uploaded to an analysis server.
Performance: Speed depends on browser, GPU support and available memory. ONNX is not always faster; the first run also includes model downloads.
When to use it: Quick analysis with no installation, including the GitHub Pages version.

Local Python

Getting started: Install Python dependencies and FFmpeg, then start the local FastAPI service. Model weights download on first use.
How it runs: FFmpeg decodes audio. Beat This! and S-KEY use PyTorch. htdemucs and GAME Small/Large run on CPU with ONNX Runtime; PyTorch handles htdemucs STFT transforms only. The separation process exits before GAME Large starts.
Audio processing: The browser sends audio over HTTP to FastAPI on this computer (127.0.0.1:8765), which returns results and a MIDI download.
Performance: Speed depends on your CPU and available memory. It avoids browser runtime constraints but needs a running local service.
When to use it: A local Python workflow, or an alternative when browser inference cannot run. See the setup guide in the engine selector above.

How are average BPM and variable tempo MIDI calculated?

Average BPM uses the number of beat intervals and the time from the first to the last beat. With at least five beats and a deviation of more than 30 ms from a constant grid, the result is marked as variable tempo. MIDI tempo changes follow each beat interval and preserve the initial beat offset. Import it at the same start time as the audio, with tempo-map import enabled in your DAW. The adjacent tempo extends before the first and after the last beat. Jitter or missed beats may also appear as tempo changes. An uncertain meter is omitted and does not by itself indicate variable tempo. When evidence is insufficient, a single average tempo is exported; insufficient or invalid beats, or tempos outside the MIDI range, return -1. Each detected beat is treated as a quarter note, so half-time and double-time ambiguity remain possible.

What does the S-KEY result mean?

S-KEY estimates the most likely tonic and major/minor mode for the whole clip, not individual chords or key changes over time. At least 3 seconds of audible audio are required. Relative keys, percussion with little melody, and modulating music may be ambiguous. This is a model estimate, not a human annotation. MIDI includes tempo, detected meter and available Lead Vocal notes; the key is displayed on this page.

Found a problem?

Report bugs or share suggestions on GitHub Issues. Include your browser, operating system and analysis engine to help us investigate.

Report an issue

Melody in your ears,notes in your hands.

01Your audio

Drop your music here

How the browser communicates with Python

1. Prepare your project folder

2. Install Python dependencies once

3. Start the local service

4. Return here and analyze

Still unable to connect?

02Analysis results

Detected vocal range

Add your audio

Discover the rhythm

Back to creating

Meet the models

Beat This!

S-KEY

GAME Small / Large

htdemucs

ONNX vs. PyTorch: model differences

Browser ONNX · Recommended

Local Python

Found a problem?

Melody in your ears,
notes in your hands.