Download video via RapidAPI (own-CDN) -> faster-whisper transcript + local screenshots -> delete; retire proxies

pipeline/asr.py

@@ -0,0 +1,60 @@
+"""Local ASR transcription with faster-whisper, run on the downloaded video file.
+
+Used when the pipeline downloads the source video (via the RapidAPI fast-downloader):
+faster-whisper reads the media directly (PyAV decodes the audio track), so no separate
+audio extraction is needed. Returns ``[{start, end, text}]`` segments — the same shape the
+caption-based path produces — so the rest of the pipeline is unchanged.
+
+Model size / device are configurable via env: ``WHISPER_MODEL`` (default ``base``),
+``WHISPER_DEVICE`` (default ``cpu``), ``WHISPER_COMPUTE`` (default ``int8`` — fastest on
+CPU). ``base``/``int8`` is a sensible CPU default; use ``small``/``medium`` (or a GPU) for
+higher accuracy.
+"""
+from __future__ import annotations
+
+import os
+from functools import lru_cache
+
+
+class ASRError(RuntimeError):
+    """Raised when local transcription fails."""
+
+
+@lru_cache(maxsize=1)
+def _model():
+    from faster_whisper import WhisperModel
+
+    size = os.environ.get("WHISPER_MODEL", "base")
+    device = os.environ.get("WHISPER_DEVICE", "cpu")
+    compute = os.environ.get("WHISPER_COMPUTE", "int8")
+    return WhisperModel(size, device=device, compute_type=compute)
+
+
+def transcribe_file(media_path: str, language: str | None = "en",
+                    progress=None) -> list[dict]:
+    """Transcribe ``media_path`` (audio or video) into ``[{start, end, text}]``.
+
+    ``language`` pins the language (faster + avoids misdetection); pass None to
+    auto-detect.
+    """
+    if not os.path.exists(media_path):
+        raise ASRError(f"media file not found: {media_path}")
+    try:
+        model = _model()
+        # No vad_filter: it depends on onnxruntime (Silero VAD) that isn't always
+        # present and, when broken, silently drops every segment. beam_size=1 for speed.
+        segments, info = model.transcribe(media_path, language=language, beam_size=1)
+    except Exception as exc:
+        raise ASRError(f"faster-whisper failed: {type(exc).__name__}: {exc}") from exc
+
+    total = float(getattr(info, "duration", 0) or 0)
+    out: list[dict] = []
+    for s in segments:  # generator — transcription happens as we iterate
+        text = (s.text or "").strip()
+        if text:
+            out.append({"start": float(s.start), "end": float(s.end), "text": text})
+        if progress and total:
+            progress(min(1.0, (s.end or 0) / total), desc="Transcribing (faster-whisper)")
+    if not out:
+        raise ASRError("faster-whisper produced no speech segments.")
+    return out