Audio to MIDI offline:
what to download first
Audio to MIDI runs locally in this converter, but our November 18 in-app-browser test failed to reload it offline after an online visit; that session did not establish a usable offline start. offline-launch-observation.json · loader-excerpts.txt
The original static file inspection is retained below, alongside one subsequent browser-enforced offline reload that failed before any audio conversion. offline-launch-observation.json
Observed offline reload: failed before the tool started
We opened the public tool online without selecting audio, then enabled Chromium tab-scoped offline mode and reloaded once without clearing or disabling the existing browser cache. offline-launch-observation.json
That reload returned ERR_INTERNET_DISCONNECTED before the converter page started, so no local model or cached inference runtime actually ran after the reload. offline-launch-observation.json
Before disconnecting, the window Performance list showed the page scripts but no converter worker, model or WASM request, which is an observed window list rather than an exhaustive worker or persistent-cache inventory. offline-launch-observation.json
This single in-app-browser result does not identify a missing cache entry or test an already warmed model, a restarted browser or other browsers; do not treat one online visit as verified offline preparation. offline-launch-observation.json
| Stage | Observed result |
|---|---|
| Online tool visit | Loaded; no audio selected or model prepared by this test |
| Offline reload | ERR_INTERNET_DISCONNECTED; converter document did not start |
| Model, inference and MIDI output | Not reached |
| Warm-model use or another browser | Not tested |
2. Page loaded and model ready are separate states
The current page script creates a separate worker from /assets/converter-v2/transcribe-worker.js inside its local-worker setup path. loader-excerpts.txt · bootstrap-audit.json
Inside the model-loader function, model.json and group1-shard1of1.bin are fetched as separate resources rather than arriving as one self-contained page. loader-excerpts.txt · asset-manifest.csv
That same function initializes the WASM backend before making the model JSON and weight requests, separating backend preparation from model-file loading in the inspected code. loader-excerpts.txt
The inspected model manifest names one weight path, group1-shard1of1.bin, so keeping only the JSON file omits the binary resource that the loader requests. bootstrap-audit.json
3. A concrete file inventory for this build
The served model JSON is 174,537 bytes and the weight file is 742,392 bytes, for a combined 916,929 raw file bytes. bootstrap-audit.json · asset-manifest.csv
The worker bundle is 1,253,307 bytes, larger than those two model files combined, so the model’s size alone understates this inspected code-and-model inventory. bootstrap-audit.json · asset-manifest.csv
The inventory contains base, SIMD and threaded-SIMD WASM files of 275,675, 377,757 and 386,833 bytes respectively, without recording which variant a particular browser chooses. bootstrap-audit.json · asset-manifest.csv
Adding the worker, model pair and one listed WASM variant gives 2,445,911–2,557,069 raw file bytes; these conditional sums exclude the page shell, compression, cache hits and actual network overhead. bootstrap-audit.json
| Runtime resource | Raw bytes | Public HTTP | Observed Cache-Control |
|---|---|---|---|
| converter-v2/transcribe-worker.js | 1253307 | 200 | public, max-age=14400, must-revalidate |
| model/model.json | 174537 | 200 | public, max-age=0, must-revalidate |
| model/group1-shard1of1.bin | 742392 | 200 | public, max-age=14400, must-revalidate |
| converter-v2/tfjs-backend-wasm.wasm | 275675 | 200 | public, max-age=0, must-revalidate |
| converter-v2/tfjs-backend-wasm-simd.wasm | 377757 | 200 | public, max-age=0, must-revalidate |
| converter-v2/tfjs-backend-wasm-threaded-simd.wasm | 386833 | 200 | public, max-age=0, must-revalidate |
4. Memory reuse and persisted files are different checks
The observed model-JSON response used max-age=0 with must-revalidate, while the weights and worker responses used max-age=14400 with must-revalidate. bootstrap-audit.json · asset-manifest.csv
The model loader retains its model-data promise in the worker-module variable av and clears that variable on a loading error; this is a code-level memory reuse mechanism, not a declared persistent-cache guarantee. loader-excerpts.txt · bootstrap-audit.json
The five named source files had zero direct serviceWorker.register and caches.open text matches in this bounded scan, which does not rule out other dynamic behavior or prove an offline failure. source-scan.json · bootstrap-audit.json
5. Locate the loading branch before retrying the recording
The source map separates a worker error callback from two non-OK response branches that name the model JSON and binary weights, giving different dependency locations to inspect. load-failure-map.json · loader-excerpts.txt
A rejected fetch can surface a network exception before a custom response-status message, so those strings are source branches rather than observed offline-error screenshots. load-failure-map.json
The source inspection and this failed reload do not establish whether a previously initialized model or another browser session would work without a connection. offline-launch-observation.json
| Source condition | Source message / branch | Dependency |
|---|---|---|
| worker.onerror | Local processing worker could not start | transcribe-worker.js / runtime environment |
| model JSON Response.ok=false | Local note model could not load. | model/model.json |
| weights Response.ok=false | Local note weights could not load. | model/group1-shard1of1.bin |
6. What publisher documentation tells you to prepare
NeuralNote’s current installation documentation says transcription runs locally but the model is not included, making the app installation and its first model download separate preparation steps. documentation-evidence.json · NeuralNote
MAGDA instead documents its Basic Pitch transcription model as bundled with the app, while describing separate first-use downloads for its Demucs and Spleeter stem models. documentation-evidence.json · stems
Basic Pitch documents a command-line workflow using local input and output paths, but its package, runtime and model availability still need to be prepared for the installation you intend to use. documentation-evidence.json · basic-pitch
Other reviewed browser publishers already describe after-load offline operation, so offline support is not presented here as an exclusive feature or a new transcription-quality advantage. documentation-evidence.json · · audio-to-midi-converter
| Publisher documentation | Preparation distinction | Status here |
|---|---|---|
| NeuralNote | App installation plus separate model download | Documentation only; not installed |
| MAGDA | Bundled transcription model; separate stem models | Documentation only; not installed |
| Basic Pitch | Local CLI plus available runtime/model files | Documentation only; not executed |
7. Prepare for the lifecycle you actually need
MDN explains that applications manage Cache entries and a browser can remove origin Cache storage, so even a stated caching mechanism is not evidence that files remain available on your device. documentation-evidence.json · Cache
Before relying on a disconnected session, obtain the app and selected model while connected and check the exact launch lifecycle that matters for your own setup. offline-checklist.md · documentation-evidence.json
The downloadable worksheet distinguishes an open page, a fresh page load and a browser restart instead of marking them all passed after one online visit. offline-checklist.md
One offline reload was attempted in the follow-up, but no new recording, audio conversion or model execution took place, so this is startup evidence rather than pitch-accuracy evidence. offline-launch-observation.json
The included auditor reads public files and source text, with an optional anonymous HTTP-file check, and never evaluates the transcription worker or imports its model. audit.mjs · README.md
The dated manifest describes this build rather than every converter, and its byte sums are not measured transfer costs, memory requirements or a runtime benchmark. bootstrap-audit.json · research-review.json
8. Choose the relevant next step
If the next problem is a file-format or note-quality question, use the linked format notes and dated quality study rather than treating it as an offline-loading diagnosis. offline-checklist.md
Use the browser converter while connected when you need a note draft; for a disconnected session, prepare and verify the exact installation you plan to rely on. offline-checklist.md