Add Faster Whisper UDF usage pattern for audio transcription

[Copilot]

Adds a usage pattern demonstrating how to run ML models like Faster Whisper using Daft's class-based UDF system.

Changes

• New file: usage_patterns/udfs/udf_faster_whisper.py - Complete example showing:

  • @daft.cls() decorator for model initialization once, reused across rows
  • @daft.method() with custom DataType schemas for structured outputs
  • daft.File handling for audio inputs with VAD transcription

• README.md: Added new pattern to docs, fixed incorrect paths for existing UDF examples

• Makefile: Added to test-udfs target

Example

@daft.cls()
class FasterWhisperTranscriber:
    def __init__(self, model="distil-large-v3", compute_type="float32", device="auto"):
        self.whisper_model = WhisperModel(model, compute_type=compute_type, device=device)
        self.pipe = BatchedInferencePipeline(self.whisper_model)

    @daft.method(return_dtype=TranscriptionResult)
    def transcribe(self, audio_file: daft.File, batch_size: int = 16):
        with audio_file.to_tempfile() as tmp:
            segments_iter, info = self.pipe.transcribe(str(tmp.name), vad_filter=True, ...)
            return {"transcript": text, "segments": segments, "info": asdict(info)}

# Usage
transcriber = FasterWhisperTranscriber()
df = (
    daft.from_glob_path("hf://datasets/Eventual-Inc/sample-files/audio/*.mp3")
    .with_column("audio_file", file(col("path")))
    .with_column("result", transcriber.transcribe(col("audio_file")))
    .select("path", unnest(col("result")))
)

Original prompt

Build me a usage pattern for running this model on daft udfs

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.