CREPE-pitch-LiteRT

Measured on device (edge-compat): Raspberry Pi 5 · LiteRT 2.2.0.dev20260804 · CPU/XNNPACK, 4 threads · 49.5 ms p50 (2026-08-31); browser · Chromium 151 on M4 Max · LiteRT.js 2.5.3 · WebGPU · 6.04 ms p50 · output matches CPU (2026-08-11). Record: https://github.com/john-rocky/edge-compat/blob/main/cards/crepe-pitch/CARD.md

CREPE — LiteRT (on-device pitch detection / real-time tuner, fully-GPU)

CREPE monophonic pitch (f0) estimation, converted to LiteRT and running fully on the CompiledModel GPU (ML Drift) on Android. A 1024-sample (16 kHz) window → activations over 360 pitch bins (20 cents each, ~C1–B7); the host decodes them to a frequency and the nearest musical note. Drives a real-time mic tuner (note + cents flat/sharp).

frame[1,1024] (16 kHz, per-frame zero-mean/unit-var) →[GPU CNN]→ activations[1,360] →[host]→ Hz → note

On-device (Pixel 8a, Tensor G3 — verified)

How it converts — the cleanest in the zoo, zero patches

The whole network is a pure CNN: 6× {zero-pad → Conv2d → ReLU → BatchNorm → MaxPool} + permute/reshape (≤4D) + Linear + sigmoid. Converted directly with litert-torch — no rewrites:

• No banned ops — the asymmetric "same" padding is a constant zero-pad → native PAD (not GATHER); no GELU / TransposeConv / dilated conv; the head permute(0,2,1,3).reshape stays ≤4D.
• No fp16-on-Mali wall — per-frame zero-mean/unit-var normalization keeps activations ~O(1).

op-check: banned NONE, >4D 0; fp16 corr 1.000000.

Preprocessing & decode (host-side)

Mono 16 kHz. Frame into 1024-sample windows; per frame subtract the mean and divide by the std. Decode the 360 activations: peak bin ± 4, activation-weighted average → cents = 20·bin + 1997.3794… → Hz = 10·2^(cents/1200); the peak activation is the confidence. Nearest note: midi = 69 + 12·log2(Hz/440).

Minimal usage

Android (Kotlin, CompiledModel GPU)

val model = CompiledModel.create(context.assets, "crepe_full_fp16.tflite",
    CompiledModel.Options(Accelerator.GPU), null)
val inputs = model.createInputBuffers(); val outputs = model.createOutputBuffers()
inputs[0].writeFloat(frame)           // [1,1024] 16 kHz window, per-frame zero-mean/unit-var
model.run(inputs, outputs)
val act = outputs[0].readFloat()      // [360] pitch-bin activations -> host decode to Hz

Python (desktop verification)

import numpy as np, soundfile as sf
from ai_edge_litert.interpreter import Interpreter

wav, _ = sf.read("note_16k.wav", dtype="float32")               # mono 16 kHz
f = wav[:1024].copy()
f = (f - f.mean()) / max(f.std(), 1e-10)                        # per-frame normalize

it = Interpreter(model_path="crepe_full_fp16.tflite"); it.allocate_tensors()
it.set_tensor(it.get_input_details()[0]["index"], f[None]); it.invoke()
act = it.get_tensor(it.get_output_details()[0]["index"])[0]     # [360]

c = int(act.argmax()); s, e = max(0, c - 4), min(360, c + 5)    # torchcrepe weighted_argmax
w, b = act[s:e], np.arange(s, e)
cents = 20.0 * (w * b).sum() / w.sum() + 1997.3794084376191
hz = 10.0 * 2 ** (cents / 1200.0)
midi = 69 + 12 * np.log2(hz / 440.0)
names = ["C", "C#", "D", "D#", "E", "F", "F#", "G", "G#", "A", "A#", "B"]
n = int(round(midi))
print(f"{hz:.1f} Hz  {names[n % 12]}{n // 12 - 1} {round((midi - n) * 100):+d} cents  conf {act[c]:.2f}")

Files

Performance

Measured on a Pixel 8a (Tensor G3, Android 16) with the standard TFLite benchmark_model tool — 10 warm-up runs then 50 timed runs, reported as the tool's mean.

The two GPU rows are different runtimes, not a contradiction. The LITERT_CL figure is the one recorded when this model shipped, taken through LiteRT's own CompiledModel accelerator — the path the Kotlin sample app and the LiteRT API use. The TfLiteGpuDelegateV2 figure is the classic TFLite OpenCL delegate, measured with a tool anyone can download and re-run. They agree on how much of the graph the GPU takes; they disagree on speed, and the classic delegate is the slower of the two here. Read the TfLiteGpuDelegateV2 row as a reproducible floor, not as this model's speed on LiteRT.

XNNPACK declines these fp16 graphs — it reports failed to delegate DEPTHWISE_CONV_2D and then fails to allocate tensors — so there is no usable CPU number. Disabling XNNPACK falls back to reference kernels, which measured about 20× slower than the GPU on models of this size and would not represent CPU inference anyone would ship.

Snapdragon NPU (Hexagon)

The NPU compiles this graph and then fails to run it: LiteRtException: Failed to invoke the compiled model. The GPU row below is the only S26 figure for it. A clean compile is not evidence that a model runs.

Measured on a Samsung Galaxy S26 (Snapdragon 8 Elite Gen 5 / SM8850, Hexagon v81, Android 16), LiteRT CompiledModel 2.2.0, one accelerator per process, 5 warm-up runs then N=50 timed runs, median reported. The run held thermal status NONE throughout. Headroom 0.62, where 1.0 is the throttling threshold.

Recipes for both accelerators: NPU · GPU.

Raspberry Pi 5 (CPU)

Measured on a Raspberry Pi 5 Model B Rev 1.1 (8 GB, Raspberry Pi OS 64-bit) with the LiteRT benchmark_model tool from litert-cli-nightly 0.2.0.dev20260805: CPU inference (XNNPACK, 4 threads), 3 invocations per file of 10 warm-up plus 50 timed runs (the tool caps a phase at 150 s, so very slow graphs run fewer — the Runs column is the actual timed total). The latency is the median across invocations; the spread is the min–max over all timed runs. No thermal throttling occurred during these runs (vcgencmd get_throttled stayed 0x0).

License

MIT. Upstream: marl/crepe (Kim, Salamon, Li, Bello — "CREPE: A Convolutional Representation for Pitch Estimation", ICASSP 2018); PyTorch weights via torchcrepe (MIT).

Citation

@inproceedings{kim2018crepe,
  title={CREPE: A Convolutional Representation for Pitch Estimation},
  author={Kim, Jong Wook and Salamon, Justin and Li, Peter and Bello, Juan Pablo},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2018}
}