arraypress commited on
Commit
4f74b09
·
verified ·
1 Parent(s): f5949c9

CREPE full and tiny for Core AI

Browse files
.gitattributes CHANGED
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tune-crepe-full-float32.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
37
+ tune-crepe-tiny-float32.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ tags:
4
+ - audio
5
+ - pitch-estimation
6
+ - crepe
7
+ - core-ai
8
+ - apple-silicon
9
+ - macos
10
+ library_name: swift-pitch-tracker
11
+ ---
12
+
13
+ # tune-crepe — CREPE for Core AI (macOS 27)
14
+
15
+ CREPE (Kim, Salamon, Li and Bello, "CREPE: A Convolutional Representation for Pitch
16
+ Estimation", ICASSP 2018; MIT) exported to Apple Core AI `.aimodel` assets for on-device pitch
17
+ tracking. The weights are the original Keras ones via [torchcrepe](https://github.com/maxrmorrison/torchcrepe)
18
+ (Max Morrison, MIT). Read by [swift-pitch-tracker](https://github.com/arraypress/swift-pitch-tracker)
19
+ and the [`tune`](https://github.com/arraypress/swift-tune-cli) command-line tool.
20
+
21
+ ## Files
22
+
23
+ | Path | What |
24
+ |---|---|
25
+ | `tune-crepe-full-float32.aimodel/` | The paper's model, 22 M parameters, 89 MB. `frames [256, 1024]` → `activations [256, 360]`. |
26
+ | `tune-crepe-tiny-float32.aimodel/` | The small model, 0.5 M parameters, 2 MB. Same shapes. |
27
+
28
+ Input: 1024-sample frames of 16 kHz audio, each normalised to zero mean and unit (biased)
29
+ standard deviation, as `crepe.core.get_activation` builds them (centre zero-pad 512, hop 160).
30
+ Output: sigmoid activations over 360 bins 20 cents apart from 1997.38 cents above 10 Hz.
31
+ Decoding (local average around the argmax, or Viterbi) is the host's, reproducing `marl/crepe`.
32
+
33
+ ## Fidelity
34
+
35
+ The exported network matches the original Keras `crepe` at 145–151 dB PSNR on the activations
36
+ (argmax identical); on Core AI it runs at 141–145 dB (full) and 145 dB (tiny) against
37
+ torchcrepe; whole clips decode to within 0.001 cents of `crepe.predict`.
38
+
39
+ ## Use
40
+
41
+ ```sh
42
+ hf download arraypress/tune-crepe --local-dir models
43
+ tune model install models
44
+ tune bass_01.wav # E1 −4 cents
45
+ ```
46
+
47
+ Requires macOS 27 (Core AI) on Apple silicon. Converted with coreai-torch 0.4.2 / torch 2.13.
tune-crepe-full-float32.aimodel/main.hash ADDED
@@ -0,0 +1 @@
 
 
1
+ 0�Ivu�iF���Q���I�ּg��s"���z�
tune-crepe-full-float32.aimodel/main.mlirb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:30b2497675d96946eef5ca518c97ea050f49afd6bc6780eb7322c51ed2c67a89
3
+ size 88991562
tune-crepe-full-float32.aimodel/metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "assetVersion" : "2.0",
3
+ "producer" : "coreai-core 1.0.0b2",
4
+ "author" : "Kim, Salamon, Li, Bello — CREPE (ISMIR 2018); PyTorch weights via torchcrepe (Max Morrison); Core AI export by PitchTracker",
5
+ "license" : "MIT",
6
+ "description" : "CREPE full: [256, 1024] normalised 16 kHz frames → [256, 360] sigmoid activations over 20-cent bins from 1997.38 cents (32.7 Hz). Framing and decoding in the host as marl\/crepe does them.",
7
+ "creationDate" : "20260924T120844Z"
8
+ }
tune-crepe-tiny-float32.aimodel/main.hash ADDED
Binary file (32 Bytes). View file
 
tune-crepe-tiny-float32.aimodel/main.mlirb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:20f5fd004fb750890a0bf5b1eb70263a298d6479000562e6f7149b6eab26150b
3
+ size 1961870
tune-crepe-tiny-float32.aimodel/metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "license" : "MIT",
3
+ "author" : "Kim, Salamon, Li, Bello — CREPE (ISMIR 2018); PyTorch weights via torchcrepe (Max Morrison); Core AI export by PitchTracker",
4
+ "assetVersion" : "2.0",
5
+ "creationDate" : "20260924T120849Z",
6
+ "description" : "CREPE tiny: [256, 1024] normalised 16 kHz frames → [256, 360] sigmoid activations over 20-cent bins from 1997.38 cents (32.7 Hz). Framing and decoding in the host as marl\/crepe does them.",
7
+ "producer" : "coreai-core 1.0.0b2"
8
+ }