Use int8 compute_type for WhisperX on non-CUDA devices

[pyjeebz]

ExtractWordsFromAudio._get_transcript_from_audio hardcodes compute_type = "float16" for the WhisperX subprocess regardless of device (tribev2/eventstransforms.py:107-108):

device = "cuda" if torch.cuda.is_available() else "cpu"
compute_type = "float16"

faster-whisper (via ctranslate2) refuses float16 on CPU and raises:

ValueError: Requested float16 compute type, but the target device or backend do not support efficient float16 computation.

Anyone without an NVIDIA GPU hits this when calling get_events_dataframe() on text or audio inputs (text goes through TTS → WhisperX too), which blocks model.predict() end-to-end on CPU-only environments.

The fix mirrors the device selection one line up — keep float16 on CUDA, fall back to int8 on CPU. int8 is what faster-whisper recommends for CPU per their README.

Repro on a CPU-only box:

from tribev2 import TribeModel
m = TribeModel.from_pretrained("facebook/tribev2")
m.get_events_dataframe(text_path="some_text.txt")  # raises before this patch

CUDA users are unaffected. Verified locally on WSL2 / torch==2.6.0+cpu: WhisperX large-v3 returned the expected word timings, and a full text-mode model.predict() completed end-to-end (output shape (n_timesteps, 20484) on fsaverage5, matching the model card). I haven't run the [test] extra's pytest suite — happy to if useful. CLA signed.

[Copilot]

Pull request overview

Fixes CPU-only execution of WhisperX transcription by selecting a supported compute_type when CUDA isn’t available, unblocking get_events_dataframe() / end-to-end model.predict() on non-NVIDIA environments.

Changes:

• Switch WhisperX compute_type from always-float16 to float16 on CUDA and int8 on CPU.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.