refactor (untested)

ctrl/transcribe_oneoff.sh

@@ -0,0 +1,121 @@
+#!/usr/bin/env bash
+# One-off "high-quality" transcription that overrides the cached transcript
+# in the most recent run directory for the given video, then re-runs the
+# merger so the enhanced transcript is regenerated using the existing frames.
+#
+# Usage:
+#   ./ctrl/transcribe_oneoff.sh <video> [language]
+#     language: optional ISO code (es, en). Omit for auto-detect.
+#
+# What this does differently from the main pipeline:
+#   1. Reuses the existing run directory (frames cache stays put).
+#   2. Preprocesses audio: loudnorm + light denoise + speech-band filter.
+#   3. Uses whisperx large-v3 with int8 quantization (fits in ~3-4 GB GPU).
+#   4. Stricter no-speech / logprob thresholds to suppress hallucinations on
+#      silent stretches (the source of the random Arabic/etc. drift).
+#   5. Backs up the old <stem>.json before overwriting, so it is recoverable.
+#   6. Re-runs process_meeting.py without --run-whisper/--diarize so the new
+#      transcript is picked up from cache and merged with the cached frames.
+
+set -euo pipefail
+
+VIDEO="${1:?usage: $0 <video> [language]}"
+LANG="${2:-}"
+
+VENV_DIR="/home/mariano/wdir/venv/def"
+if [[ ! -f "$VENV_DIR/bin/activate" ]]; then
+    echo "ERROR: venv not found at $VENV_DIR" >&2
+    exit 1
+fi
+# shellcheck disable=SC1091
+source "$VENV_DIR/bin/activate"
+
+WHISPERX="whisperx"
+PYTHON="python"
+
+if [[ ! -f "$VIDEO" ]]; then
+    echo "ERROR: video not found: $VIDEO" >&2
+    exit 1
+fi
+
+STEM="$(basename "${VIDEO%.*}")"
+
+# Locate the most recent run dir for this video stem.
+RUN_DIR="$(ls -1dt output/*-"$STEM" 2>/dev/null | head -n1 || true)"
+if [[ -z "$RUN_DIR" || ! -d "$RUN_DIR" ]]; then
+    echo "ERROR: no existing run dir found under output/ for stem '$STEM'." >&2
+    echo "       Run process_meeting.py at least once first to extract frames." >&2
+    exit 1
+fi
+echo "==> Using run dir: $RUN_DIR"
+
+TRANSCRIPT_JSON="$RUN_DIR/${STEM}.json"
+
+# Back up the old transcript before overwrite.
+if [[ -f "$TRANSCRIPT_JSON" ]]; then
+    BACKUP="$TRANSCRIPT_JSON.bak.$(date +%Y%m%d-%H%M%S)"
+    cp "$TRANSCRIPT_JSON" "$BACKUP"
+    echo "==> Backed up old transcript → $BACKUP"
+fi
+
+# No audio preprocessing: previous attempts with afftdn/loudnorm caused VAD
+# to drop the bulk of the meeting after ~20min. Feed the raw video directly;
+# whisperx will extract audio internally.
+INPUT_AUDIO="$VIDEO"
+
+# cuDNN libs for whisperx (mirrors what process_meeting.py does).
+SITE_PKGS="$(python -c 'import site; print(site.getsitepackages()[0])')"
+CUDNN_LIB="$SITE_PKGS/nvidia/cudnn/lib"
+if [[ -d "$CUDNN_LIB" ]]; then
+    export LD_LIBRARY_PATH="$CUDNN_LIB:${LD_LIBRARY_PATH:-}"
+fi
+
+# whisperx writes <input_basename>.json into --output_dir.
+# Our input basename is "${STEM}_clean", so we redirect to a temp dir and
+# move the result into place under the canonical name.
+TX_TMP="$(mktemp -d)"
+trap 'rm -rf "$TX_TMP"' EXIT
+
+CMD=(
+    "$WHISPERX" "$INPUT_AUDIO"
+    --model large-v3
+    --compute_type int8
+    --batch_size 4
+    --output_format json
+    --output_dir "$TX_TMP"
+    --diarize
+)
+
+if [[ -n "$LANG" ]]; then
+    echo "==> Forcing language: $LANG"
+    CMD+=(--language "$LANG")
+else
+    echo "==> Auto-detecting language"
+fi
+
+if [[ -n "${HF_TOKEN:-}" ]]; then
+    CMD+=(--hf_token "$HF_TOKEN")
+fi
+
+echo "==> Running: ${CMD[*]}"
+"${CMD[@]}"
+
+# Move whisperx output into place under the canonical name expected by the cache.
+# whisperx names output by the input basename (without extension).
+NEW_JSON="$TX_TMP/${STEM}.json"
+if [[ ! -f "$NEW_JSON" ]]; then
+    echo "ERROR: expected whisperx output not found: $NEW_JSON" >&2
+    exit 1
+fi
+mv "$NEW_JSON" "$TRANSCRIPT_JSON"
+echo "==> Wrote new transcript → $TRANSCRIPT_JSON"
+
+echo
+echo "==> Re-running merger to regenerate enhanced transcript with cached frames"
+"$PYTHON" process_meeting.py "$VIDEO" \
+    --embed-images \
+    --scene-detection \
+    --scene-threshold 10
+
+echo
+echo "==> Done."