DRTR-J commited on
Commit
525b00a
·
verified ·
1 Parent(s): 76ed5cb

Update whisper-large-v3-turbo-coreml-fp16/README.md

Browse files
whisper-large-v3-turbo-coreml-fp16/README.md CHANGED
@@ -6,7 +6,7 @@ Core ML export of `openai/whisper-large-v3-turbo` tuned for Apple Silicon. This
6
  - `DecoderFull.mlpackage` – full-context fallback
7
  - `DecoderStateful.mlpackage` – experimental MLState variant (requires macOS 15+)
8
 
9
- Tokenizers, mel filters, and metadata are included so the bundle can be dropped directly into a Tauri, Swift, or whisper.cpp/Core ML integration.
10
 
11
  ## Contents
12
 
@@ -32,22 +32,23 @@ preprocessor_config.json
32
  ## Quick Start (Swift)
33
 
34
  ```swift
35
- let bundleURL = URL(fileURLWithPath: "../Models/whisper-large-v3-turbo-coreml-fp16")
36
- let whisper = try WhisperCoreML(
37
- encoderURL: bundleURL.appendingPathComponent("Encoder.mlpackage"),
38
- decoderURL: bundleURL.appendingPathComponent("DecoderWithCache.mlpackage"),
39
- tokenizerDirectory: bundleURL
40
- )
41
- let result = try await whisper.transcribe(audioURL: inputURL)
42
- print(result.text)
43
  ```
44
 
45
- ## Quick Start (CLI sandbox)
46
 
47
  ```bash
48
- ./tauri-coreml-sandbox/src-tauri/target/release/coreml_sandbox \
49
- artifacts/hf-upload/whisper-large-v3-turbo-coreml-fp16 \
50
- tests/audio/2p_argument.wav
 
51
  ```
52
 
53
  ## Performance Snapshot
 
6
  - `DecoderFull.mlpackage` – full-context fallback
7
  - `DecoderStateful.mlpackage` – experimental MLState variant (requires macOS 15+)
8
 
9
+ Tokenizers, mel filters, and metadata are included so the bundle can be dropped directly into any Core ML-driven Whisper integration.
10
 
11
  ## Contents
12
 
 
32
  ## Quick Start (Swift)
33
 
34
  ```swift
35
+ let bundleURL = URL(fileURLWithPath: "/path/to/whisper-large-v3-turbo-coreml-fp16")
36
+ let encoderURL = bundleURL.appendingPathComponent("Encoder.mlpackage")
37
+ let decoderURL = bundleURL.appendingPathComponent("DecoderWithCache.mlpackage")
38
+
39
+ let encoder = try MLModel(contentsOf: encoderURL)
40
+ let decoder = try MLModel(contentsOf: decoderURL)
41
+ // Plug these into your transcription pipeline together with the tokenizer assets
42
+ // located alongside the mlpackages.
43
  ```
44
 
45
+ ## Quick Start (Python parity check)
46
 
47
  ```bash
48
+ python3 scripts/verify_coreml_whisper.py \
49
+ --audio /path/to/audio.wav \
50
+ --coreml-dir /path/to/whisper-large-v3-turbo-coreml-fp16 \
51
+ --max-new-tokens 64
52
  ```
53
 
54
  ## Performance Snapshot