add examples/whisper.youtube: YouTube → speech-to-text pipeline

- yt_transcribe.py: yt-dlp + ffmpeg + whisper-cli (large-v3) wrapper
- parakeet_transcribe.py: alternative NVIDIA Parakeet ASR backend
- parakeet: shell launcher for parakeet via local venv
- transcriptions/q1CluXct_VI.txt: sample output (8h audiobook)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

examples/whisper.youtube/.gitignore

@@ -0,0 +1,4 @@
+videos/
+.venv-parakeet/
+__pycache__/
+*.pyc

examples/whisper.youtube/parakeet

@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+# Parakeet V3 transcription launcher.
+# Usage:
+#   ./parakeet <wav-file>
+#   ./parakeet <youtube-url>
+set -euo pipefail
+DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+exec "$DIR/.venv-parakeet/bin/python" "$DIR/parakeet_transcribe.py" "$@"

examples/whisper.youtube/parakeet_transcribe.py

@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+"""Transcribe audio with NVIDIA Parakeet-tdt-0.6b-v3 (multilingual)."""
+
+import os
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+MODEL_NAME = "nvidia/parakeet-tdt-0.6b-v3"
+
+
+def download_audio(url: str, out_path: str):
+    subprocess.run([
+        "yt-dlp", "-x", "--audio-format", "wav",
+        "-o", out_path, url,
+    ], check=True)
+
+
+def convert_wav(src: str, dst: str):
+    subprocess.run([
+        "ffmpeg", "-y", "-i", src,
+        "-ar", "16000", "-ac", "1", "-f", "wav", dst,
+    ], check=True, capture_output=True)
+
+
+def transcribe(wav_path: str, long_audio: bool = True) -> str:
+    import nemo.collections.asr as nemo_asr
+
+    print(f"Loading {MODEL_NAME} ...", flush=True)
+    asr = nemo_asr.models.ASRModel.from_pretrained(MODEL_NAME)
+    asr = asr.cuda() if __import__("torch").cuda.is_available() else asr
+
+    if long_audio:
+        asr.change_attention_model(
+            self_attention_model="rel_pos_local_attn",
+            att_context_size=[256, 256],
+        )
+
+    print(f"Transcribing {wav_path} ...", flush=True)
+    out = asr.transcribe([wav_path], timestamps=False)
+    return out[0].text
+
+
+def main():
+    if len(sys.argv) < 2:
+        print(f"Usage: {sys.argv[0]} <wav-file-or-youtube-url>")
+        sys.exit(1)
+
+    src = sys.argv[1]
+
+    if src.startswith(("http://", "https://", "youtu", "www.")):
+        with tempfile.TemporaryDirectory(prefix="parakeet-") as tmp:
+            raw = os.path.join(tmp, "audio.%(ext)s")
+            wav = os.path.join(tmp, "audio_16k.wav")
+            print(f"Downloading {src} ...", flush=True)
+            download_audio(src, raw)
+            downloaded = [f for f in os.listdir(tmp) if f.startswith("audio.")]
+            print("Converting to 16kHz WAV ...", flush=True)
+            convert_wav(os.path.join(tmp, downloaded[0]), wav)
+            text = transcribe(wav)
+    else:
+        text = transcribe(src)
+
+    print("\n=== Transcription ===")
+    print(text)
+
+
+if __name__ == "__main__":
+    main()

examples/whisper.youtube/yt_transcribe.py

@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+"""Download audio from YouTube and transcribe with local whisper-cli."""
+
+import os
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+WHISPER_CLI = REPO_ROOT / "build" / "bin" / "whisper-cli"
+MODEL = REPO_ROOT / "models" / "ggml-large-v3.bin"
+
+
+def download_audio(url: str, out_path: str):
+    """Download audio from YouTube URL using yt-dlp."""
+    subprocess.run([
+        "yt-dlp", "-x", "--audio-format", "wav",
+        "-o", out_path, url,
+    ], check=True)
+
+
+def convert_wav(src: str, dst: str):
+    """Convert audio to 16kHz mono WAV for whisper."""
+    subprocess.run([
+        "ffmpeg", "-y", "-i", src,
+        "-ar", "16000", "-ac", "1", "-f", "wav", dst,
+    ], check=True, capture_output=True)
+
+
+def transcribe(wav_path: str) -> str:
+    """Run whisper-cli on a WAV file and return text."""
+    result = subprocess.run([
+        str(WHISPER_CLI), "-m", str(MODEL),
+        "-f", wav_path, "-nt", "-np", "-l", "auto",
+    ], capture_output=True, text=True)
+    if result.returncode != 0:
+        print(f"whisper-cli error: {result.stderr[:500]}", file=sys.stderr)
+        sys.exit(1)
+    return result.stdout.strip().replace("[BLANK_AUDIO]", "").strip()
+
+
+def main():
+    if len(sys.argv) < 2:
+        print(f"Usage: {sys.argv[0]} <youtube-url>")
+        sys.exit(1)
+
+    url = sys.argv[1]
+
+    if not WHISPER_CLI.exists():
+        print(f"whisper-cli not found: {WHISPER_CLI}", file=sys.stderr)
+        sys.exit(1)
+    if not MODEL.exists():
+        print(f"Model not found: {MODEL}\nRun: bash models/download-ggml-model.sh base", file=sys.stderr)
+        sys.exit(1)
+
+    with tempfile.TemporaryDirectory(prefix="yt-whisper-") as tmpdir:
+        raw_audio = os.path.join(tmpdir, "audio.%(ext)s")
+        wav_16k = os.path.join(tmpdir, "audio_16k.wav")
+
+        print(f"Downloading audio from {url} ...")
+        download_audio(url, raw_audio)
+
+        # find the downloaded file (yt-dlp replaces %(ext)s)
+        downloaded = [f for f in os.listdir(tmpdir) if f.startswith("audio.")]
+        if not downloaded:
+            print("Download failed: no audio file found", file=sys.stderr)
+            sys.exit(1)
+        raw_path = os.path.join(tmpdir, downloaded[0])
+
+        print("Converting to 16kHz WAV ...")
+        convert_wav(raw_path, wav_16k)
+
+        print("Transcribing ...")
+        text = transcribe(wav_16k)
+
+        print("\n=== Transcription ===")
+        print(text)
+
+
+if __name__ == "__main__":
+    main()