osobhandClaude Opus 4.7
7e37e45df7
Wrapper that pulls audio from any yt-dlp-supported URL (YouTube, LibriVox, archive.org, podcast feeds) and converts to the 24 kHz mono 16-bit PCM format examples/audio_to_manifest ingests. Slugifies the output filename so manifest paths stay shell-safe. Prints the next-step audio_to_manifest command with all the right flags (--auto-emotion-tag --use-emotion2vec --stage audiobook), so a new user can copy-paste the printed line straight into a terminal. Requires external tools (yt-dlp, ffmpeg); install on macOS via `brew install yt-dlp ffmpeg`. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
94 lines
2.8 KiB
Bash
Executable File
94 lines
2.8 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
# Fetch audio from a URL (YouTube, archive.org, LibriVox, podcast feed,
|
|
# anything yt-dlp supports) and convert to the 24 kHz mono WAV format
|
|
# our data-prep pipeline expects.
|
|
#
|
|
# Output is suitable as direct input to `examples/audio_to_manifest`.
|
|
#
|
|
# Requires: yt-dlp, ffmpeg (install via `brew install yt-dlp ffmpeg`).
|
|
#
|
|
# Usage:
|
|
# scripts/fetch_audio.sh <url> [output_dir]
|
|
#
|
|
# Examples:
|
|
# # Single LibriVox chapter
|
|
# scripts/fetch_audio.sh \
|
|
# "https://librivox.org/.../alice-in-wonderland-chapter-1.mp3" \
|
|
# /tmp/voice_corpus
|
|
#
|
|
# # YouTube talk
|
|
# scripts/fetch_audio.sh \
|
|
# "https://www.youtube.com/watch?v=XXXXXXXXX" \
|
|
# /tmp/voice_corpus
|
|
#
|
|
# # Multiple URLs from a file (one per line)
|
|
# while read url; do
|
|
# scripts/fetch_audio.sh "$url" /tmp/voice_corpus
|
|
# done < urls.txt
|
|
#
|
|
# Tips for good training audio:
|
|
# - Single dominant speaker per clip (use diarize step if mixed)
|
|
# - At least 10-30 minutes total per persona
|
|
# - Studio-quality > phone > YouTube > heavily-compressed
|
|
# - Emotional variety beats raw duration
|
|
|
|
set -euo pipefail
|
|
|
|
if [[ $# -lt 1 ]]; then
|
|
echo "usage: $0 <url> [output_dir]" >&2
|
|
exit 1
|
|
fi
|
|
|
|
URL="$1"
|
|
OUT_DIR="${2:-./fetched_audio}"
|
|
|
|
mkdir -p "$OUT_DIR"
|
|
|
|
# yt-dlp emits the title as the filename. Strip whitespace + special
|
|
# chars in a post-rename so the manifest's `wav` paths stay shell-safe.
|
|
TMP_DIR="$(mktemp -d)"
|
|
trap "rm -rf '$TMP_DIR'" EXIT
|
|
|
|
echo "→ downloading audio from: $URL"
|
|
yt-dlp \
|
|
--extract-audio \
|
|
--audio-format wav \
|
|
--audio-quality 0 \
|
|
--no-playlist \
|
|
--output "$TMP_DIR/%(title).100s.%(ext)s" \
|
|
"$URL"
|
|
|
|
# Find what we just downloaded.
|
|
SRC_WAV="$(find "$TMP_DIR" -maxdepth 1 -name '*.wav' | head -n1)"
|
|
if [[ -z "$SRC_WAV" ]]; then
|
|
echo "error: yt-dlp produced no .wav file" >&2
|
|
exit 2
|
|
fi
|
|
|
|
# Slugify the basename: lowercase, replace non-alphanumeric with `_`,
|
|
# collapse runs, trim.
|
|
BASE="$(basename "$SRC_WAV" .wav)"
|
|
SLUG="$(echo "$BASE" \
|
|
| tr '[:upper:]' '[:lower:]' \
|
|
| sed -E 's/[^a-z0-9]+/_/g; s/^_+|_+$//g')"
|
|
DST_WAV="$OUT_DIR/${SLUG}.wav"
|
|
|
|
echo "→ converting to 24 kHz mono 16-bit PCM: $DST_WAV"
|
|
ffmpeg -y -loglevel error \
|
|
-i "$SRC_WAV" \
|
|
-ac 1 -ar 24000 -sample_fmt s16 \
|
|
"$DST_WAV"
|
|
|
|
# Print duration so the user sees what they got.
|
|
DURATION="$(ffmpeg -i "$DST_WAV" 2>&1 | grep Duration | awk '{print $2}' | tr -d ',')"
|
|
SIZE_MB="$(du -m "$DST_WAV" | awk '{print $1}')"
|
|
echo
|
|
echo "✓ wrote $DST_WAV (${DURATION}, ${SIZE_MB} MB)"
|
|
echo
|
|
echo "Next step:"
|
|
echo " cargo run -p rtx-csm --release --features metal --example audio_to_manifest -- \\"
|
|
echo " --in $DST_WAV \\"
|
|
echo " --out-dir ${OUT_DIR}_manifest \\"
|
|
echo " --wavlm-sv-weights /tmp/wavlm_sv.safetensors \\"
|
|
echo " --auto-emotion-tag --use-emotion2vec --stage audiobook"
|