The WeaveSignal contributors
16833c96ae
- tests/demo-video/: the script, designed voices, the run through the real desk and the edit that make the website's demo video (the media themselves are not in git). - The home page plays the video, with its credits. - deploy/cloudron/CHANGELOG.md: one line per entry; CloudronVersions.json rebuilt.
75 lines
4.6 KiB
Shell
Executable file
75 lines
4.6 KiB
Shell
Executable file
#!/usr/bin/env bash
|
|
# SPDX-License-Identifier: AGPL-3.0-only
|
|
# SPDX-FileCopyrightText: 2026 The WeaveSignal contributors
|
|
#
|
|
# Reads every line of tests/demo-video/script.json (the introduction, the show and the outro) in
|
|
# its character's designed voice through VoiceLab's tts-serve (127.0.0.1:7500), with the
|
|
# reference design.sh made (target/demo-video/voices/<who>.wav) and the character's seed, then
|
|
# checks it with VoiceLab's Whisper (127.0.0.1:8181).
|
|
#
|
|
# tests/demo-video/synth.sh # every line not yet made
|
|
# tests/demo-video/synth.sh l-open # these lines again (a new take)
|
|
# SEED_OFFSET=7 tests/demo-video/synth.sh l-open # a new take at the character's seed + 7
|
|
# (keep a take that passes in `reseeded`)
|
|
#
|
|
# Out, in target/demo-video/stems/:
|
|
# raw/<id>.wav the model's take, 48 kHz stereo
|
|
# <id>.wav what is played: callers coloured as a phone line (the filter of the talk-show
|
|
# fixture's fx.sh: 8 kHz, 300-3,400 Hz, a firm compressor, a little line noise;
|
|
# the desk band-limits callers again on its own, ADR-0194, which does no harm),
|
|
# every line set to -23 LUFS (the desk's leveller does the rest)
|
|
# check.tsv id, who, the words Whisper heard against the script's, and the text heard
|
|
# The text sent to the model has `say_as` respellings applied; the script keeps the real
|
|
# spelling, and `heard_as` lists what Whisper may write for a name. A line Whisper mishears is
|
|
# printed with MISHEARD: make it again with another SEED_OFFSET, or re-word it in the script.
|
|
set -euo pipefail
|
|
here=$(cd "$(dirname "$0")" && pwd)
|
|
root=$(cd "$here/../.." && pwd)
|
|
script="$here/script.json"
|
|
voices="$root/target/demo-video/voices"
|
|
out="$root/target/demo-video/stems"
|
|
TTS=${TTS:-http://127.0.0.1:7500}
|
|
STT=${STT:-http://127.0.0.1:8181}
|
|
mkdir -p "$out/raw"
|
|
PHONE="aresample=8000,highpass=f=300:poles=2,lowpass=f=3400:poles=2,acompressor=threshold=-24dB:ratio=6:attack=5:release=80:makeup=6dB,aresample=48000"
|
|
|
|
# Every line: id, who, text.
|
|
lines=$(jq -c '[.intro[], .outro, (.moments[].steps[] | (., (.background // [])[]) | select(.line) | {id: .line, who, text})] | .[]' "$script")
|
|
want=("$@")
|
|
[ -f "$out/check.tsv" ] || : > "$out/check.tsv"
|
|
|
|
. "$here/words.sh"
|
|
|
|
while IFS= read -r l; do
|
|
id=$(jq -r .id <<<"$l"); who=$(jq -r .who <<<"$l"); text=$(jq -r .text <<<"$l")
|
|
if [ ${#want[@]} -gt 0 ]; then
|
|
printf '%s\n' "${want[@]}" | grep -qx "$id" || continue
|
|
elif [ -f "$out/$id.wav" ]; then
|
|
continue
|
|
fi
|
|
seed=$(jq -r --arg w "$who" --arg i "$id" --argjson o "${SEED_OFFSET:-0}" \
|
|
'if env.SEED_OFFSET then .cast[$w].seed + $o else (.reseeded[$i] // .cast[$w].seed) end' "$script")
|
|
effect=$(jq -r --arg w "$who" '.cast[$w].effect // ""' "$script")
|
|
said=$(jq -rn --arg t "$text" --slurpfile s "$script" '$s[0].say_as as $m | $t | reduce ($m | to_entries[]) as $e (.; gsub("\\b" + $e.key + "\\b"; $e.value))')
|
|
jq -n --arg text "$said" --rawfile ref_text "$voices/$who.txt" --rawfile audio_base64 <(base64 -w0 "$voices/$who.wav") \
|
|
--argjson seed "$seed" '{text: $text, audio_base64: $audio_base64, reference_text: $ref_text, language: "en", seed: $seed}' \
|
|
| curl -fsS --max-time 300 -H 'Content-Type: application/json' --data-binary @- "$TTS/synthesize" \
|
|
| jq -r .audio_base64 | base64 -d > "$out/raw/$id.tmp.wav"
|
|
ffmpeg -v error -y -i "$out/raw/$id.tmp.wav" -ac 2 -ar 48000 "$out/raw/$id.wav"; rm -f "$out/raw/$id.tmp.wav"
|
|
if [ "$effect" = phone ]; then
|
|
ffmpeg -v error -y -i "$out/raw/$id.wav" -f lavfi -i "anoisesrc=color=pink:amplitude=0.004:r=48000" \
|
|
-filter_complex "[0:a]$PHONE,aformat=channel_layouts=stereo[v];[1:a]highpass=f=300,lowpass=f=3400,aformat=channel_layouts=stereo[n];[v][n]amix=inputs=2:duration=first:normalize=0,loudnorm=I=-23:TP=-2:LRA=11" \
|
|
-ar 48000 -ac 2 "$out/$id.wav"
|
|
else
|
|
ffmpeg -v error -y -i "$out/raw/$id.wav" -af "loudnorm=I=-23:TP=-2:LRA=11" -ar 48000 -ac 2 "$out/$id.wav"
|
|
fi
|
|
heard=$(curl -fsS --max-time 120 -F "file=@$out/$id.wav" -F response_format=json -F temperature=0 "$STT/v1/audio/transcriptions" | jq -r .text | tr '\n' ' ' | sed 's/^ *//; s/ *$//')
|
|
a=$(norm "$text"); b=$(norm "$heard")
|
|
d=$(distance "$a" "$b")
|
|
dur=$(ffprobe -v error -show_entries format=duration -of csv=p=0 "$out/$id.wav")
|
|
mark=ok; [ "$d" -gt 0 ] && mark=MISHEARD
|
|
printf '%-14s %-7s %5.2fs %-8s %s\n' "$id" "$who" "$dur" "$mark$( [ "$d" -gt 0 ] && echo " ($d)")" "$heard"
|
|
grep -v "^$id " "$out/check.tsv" > "$out/check.tsv.new" || true
|
|
printf '%s\t%s\t%s\t%s\t%s\t%s\n' "$id" "$who" "$seed" "$dur" "$d" "$heard" >> "$out/check.tsv.new"
|
|
mv "$out/check.tsv.new" "$out/check.tsv"
|
|
done <<<"$lines"
|