schism-audio/basic-pitch-coreml

Basic Pitch — Core ML

Spotify's Basic Pitch (Bittner et al., ICASSP 2022 — a ~17k-parameter instrument-agnostic model for polyphonic note transcription and multipitch estimation, audio → MIDI) for Core ML on Apple devices. BasicPitch_fp32.mlpackage is Spotify's official Core ML artifact (repo commit fa5997af, sha256-pinned, redistributed bit-identical); the fp16 variant is derived from it by recasting the CNN body while keeping the CQT frontend in FLOAT32.

Verified against the official nmp.onnx reference — the same oracle the schism-mlx MLX port is verified against (posteriorgram parity and note-event decode agreement on synthetic melodies and an E-GMD drum clip). Two variants per model:

File	Precision	Compute units	max posteriorgram diff vs nmp.onnx
BasicPitch_fp16.mlpackage	FLOAT16 body, FLOAT32 CQT frontend	ALL (ANE)	1.3e-3, decoded note events identical on tested clips
BasicPitch_fp32.mlpackage	FLOAT32 (official artifact, iOS 15+)	CPU+GPU	1.3e-6

Verified on-device-equivalently via coremltools on an M5 Max, against the reference implementation on real audio. fp16 is ANE-eligible and recommended for iPhone / iPad; fp32 is the tight-parity fallback.

Download

.mlpackage bundles must be materialized as real files — the Core ML compiler rejects the symlinks that a default snapshot_download creates in the Hugging Face cache:

from huggingface_hub import snapshot_download
path = snapshot_download("schism-audio/basic-pitch-coreml", local_dir="./basic-pitch-coreml")

(or hf download schism-audio/basic-pitch-coreml --local-dir ./basic-pitch-coreml). Swift hosts downloading files directly are unaffected.

I/O contract

• input input_2: (1, 43844, 1) float32 — one raw ~2 s window of 22050 Hz mono audio (43844 = 2·22050 − 256)
• output Identity: (1, 172, 264) float32 — pitch-contour posteriorgram (3 bins/semitone from A0)
• output Identity_1: (1, 172, 88) float32 — note posteriorgram (semitones A0–C8)
• output Identity_2: (1, 172, 88) float32 — onset posteriorgram
• frames are ~86.13 fps (22050/256); all outputs are sigmoid probabilities
• full tracks: front-pad 3840 zeros, windows every 36164 samples, drop 15 frames per side of each window's 172 output frames, concatenate, trim to int(n_samples/36164*142); frame→seconds includes a per-window offset with the reference's 0.0018 s magic alignment constant — the exact recipe (plus note-event decoding defaults) is in test_vectors_cqt.npz params and implemented in schism-mlx transcribe/basic_pitch/model.py
• caveat: skip the model for windows that are entirely zero and emit zero posteriorgram frames instead — the official artifact's normalization diverges from nmp.onnx on exact digital silence (≥ 2 s of zero samples) and can hallucinate notes; any nonzero sample in the window is fine (verified to 7e-7 on a mostly-zero window)

DSP frontend (host-side)

Unlike the other models in this org, the Core ML graph embeds the whole DSP frontend (multirate CQT, log normalization, harmonic stacking) — the host feeds raw audio windows and implements only windowing, output unwrapping, frame-time mapping and note-event decoding, all documented in the test_vectors_cqt.npz params json. The vectors also pin every internal frontend stage (CQT magnitude, normalized log-CQT, harmonic stack — float32, computed by the verified schism-mlx port, 1e-6 parity vs nmp.onnx) alongside the posteriorgrams the shipped fp32 package produces for the same window, for hosts that want a native frontend or to validate the pipeline end to end. Note-event decoding is pure post-processing (peak-picking + note segmentation); the schism-mlx numpy transliteration is verified exactly equal to basic_pitch.note_creation.

License

Apache-2.0, same as the upstream code and weights. BasicPitch_fp32.mlpackage is the unmodified official artifact from spotify/basic-pitch; BasicPitch_fp16.mlpackage is derived from it. Cite the ICASSP 2022 paper if you use this model in research. Core ML packaging and verification by schism-audio.