feat: usage examples

README.md

@@ -59,3 +59,7 @@ The following line will put it on path for the current session. If you use a vir
 ```
 export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:`python3 -c 'import os; import nvidia.cublas.lib; import nvidia.cudnn.lib; print(os.path.dirname(nvidia.cublas.lib.__file__) + ":" + os.path.dirname(nvidia.cudnn.lib.__file__))'`
 ```
+
+# Usage Guidance
+
+See [examples](./examples) for how to use the models.

examples/with_faster_whisper.py

@@ -0,0 +1,13 @@
+"""
+Example using faster-whisper with ctranslate2 backend for fast audio transcription.
+Run with:
+    pip install faster-whisper
+    python with_faster_whisper.py
+"""
+
+import faster_whisper
+model = faster_whisper.WhisperModel('ivrit-ai/whisper-large-v3-turbo-ct2')
+
+segs, _ = model.transcribe('audio.opus', language='he')
+text = ' '.join(s.text for s in segs)
+print(f'Transcribed text: {text}')

examples/with_stable_timestamps.py

@@ -0,0 +1,13 @@
+"""
+Example of Stable-ts with faster-whisper for fast and accurate transcription.
+Run with:
+    pip install -U 'stable-ts[fw]'
+    python with_stable_timestamps.py
+"""
+
+import stable_whisper
+
+model = stable_whisper.load_faster_whisper('ivrit-ai/whisper-large-v3-turbo-ct2')
+segs = model.transcribe('audio.opus', language='he') # Word level timestamps enabled by default
+for s in segs:
+    print(f'{s.start:.2f} - {s.end:.2f}: {s.text}')

examples/with_whisper_cpp.py

@@ -0,0 +1,21 @@
+"""
+Example of using whispercpp for fast and lightweight transcription.
+Download the model from Hugging Face:
+    wget https://huggingface.co/ivrit-ai/whisper-large-v3-turbo-ggml/resolve/main/ggml-model.bin
+Run with:
+    pip install pywhispercpp huggingface-hub
+    python with_whisper_cpp.py
+"""
+
+from pywhispercpp.model import Model
+from huggingface_hub import hf_hub_download
+
+
+model_path = hf_hub_download(
+    repo_id="ivrit-ai/whisper-large-v3-turbo-ggml",
+    filename="ggml-model.bin"
+)
+model = Model(model_path)
+segs = model.transcribe('audio.opus', language='he')
+text = ' '.join(segment.text for segment in segs)
+print(f'Transcribed text: {text}')

pyproject.toml

@@ -1,3 +1,11 @@
+[project]
+name = "asr-training"
+version = "0.1.0"
+description = "ASR training recipes created for ivrit.ai"
+readme = "README.md"
+requires-python = ">=3.10"
+dependencies = []
+
 [tool.black]
 line-length = 120