The trailer narrated by Old Tam: a line over every shot (welcome, survive, build, the wolves, winter, the cliffs, the Cove, play with friends), the game's sound dipping under him

tools/trailer/make_trailer.sh

@@ -28,18 +28,40 @@ OUT_AT=$(echo "$D - 2.5" | bc)
 # A slow A-minor-ish drone, left and right a little apart, echoed and soft.
 PAD="0.06*(sin(2*PI*110*t)+0.7*sin(2*PI*164.8*t)+0.5*sin(2*PI*220*t)+0.35*sin(2*PI*261.6*t)*(0.5+0.5*sin(2*PI*0.08*t)))*(0.75+0.25*sin(2*PI*0.21*t))"
 PAD_R="0.06*(sin(2*PI*110.4*t)+0.7*sin(2*PI*165.2*t)+0.5*sin(2*PI*219.6*t)+0.35*sin(2*PI*262*t)*(0.5+0.5*cos(2*PI*0.08*t)))*(0.75+0.25*cos(2*PI*0.21*t))"
-# Old Tam (0.6.2: tools/trailer/voice, made by make_trailer_voice.py), each
-# line a moment into its shot: the dawn flyover (shot 0), the wolves (3) and
-# the Cove (6). The shots' start times are in the render log.
-at() { awk -v n="$1" -v d="$2" '$1 == "[shot]" && $2 == n { printf "%d", ($3 + d) * 1000 }' build/trailer/render.log; }
+# Old Tam narrates (0.6.2: tools/trailer/voice, made by make_trailer_voice.py):
+# each line `at` seconds into its shot, by the shots' start times in the
+# render log. The game's sound and the music dip under him.
 V=tools/trailer/voice
-T_DAWN=$(at 0 2.2); T_WOLVES=$(at 3 1.0); T_COVE=$(at 6 1.2)
-[ -n "$T_DAWN" ] && [ -n "$T_WOLVES" ] && [ -n "$T_COVE" ] || { echo "no shot times in build/trailer/render.log"; exit 1; }
-ffmpeg -v error -y -i "$RAW" -f lavfi -i "aevalsrc='$PAD|$PAD_R':s=48000:d=$D" \
-	-i "$V/dawn.ogg" -i "$V/wolves.ogg" -i "$V/cove.ogg" \
+NARR=$(python3 - "$V/lines.json" build/trailer/render.log <<'EOF'
+import json, sys
+lines = json.load(open(sys.argv[1]))
+starts = {}
+for l in open(sys.argv[2]):
+    p = l.split()
+    if len(p) == 3 and p[0] == "[shot]":
+        starts[int(p[1])] = float(p[2])
+inputs, chains, labels = [], [], []
+for n, x in enumerate(lines):
+    if x["shot"] not in starts:
+        sys.exit("no start time for shot %d in the render log" % x["shot"])
+    ms = int((starts[x["shot"]] + x["at"]) * 1000)
+    inputs.append("%s/%s" % (sys.argv[1].rsplit("/", 1)[0], x["file"]))
+    chains.append("[%d:a]aresample=48000,adelay=%d|%d,volume=1.7[n%d]" % (n + 2, ms, ms, n))
+    labels.append("[n%d]" % n)
+print("\t".join(inputs))
+print(";".join(chains) + ";" + "".join(labels) + "amix=inputs=%d:duration=longest:normalize=0[v]" % len(labels))
+EOF
+) || exit 1
+NARR_IN=$(printf '%s\n' "$NARR" | head -1)
+NARR_FX=$(printf '%s\n' "$NARR" | tail -1)
+set --
+OLDIFS=$IFS; IFS='	'
+for f in $NARR_IN; do set -- "$@" -i "$f"; done
+IFS=$OLDIFS
+ffmpeg -v error -y -i "$RAW" -f lavfi -i "aevalsrc='$PAD|$PAD_R':s=48000:d=$D" "$@" \
 	-filter_complex "[1:a]lowpass=f=1800,aecho=0.8:0.7:180|340:0.35|0.25,afade=t=in:d=2.5,afade=t=out:st=$OUT_AT:d=2.5[m];[0:a]volume=0.8[g];\
-[2:a]aresample=48000,adelay=$T_DAWN|$T_DAWN,volume=1.6[v0];[3:a]aresample=48000,adelay=$T_WOLVES|$T_WOLVES,volume=1.6[v1];[4:a]aresample=48000,adelay=$T_COVE|$T_COVE,volume=1.6[v2];\
-[v0][v1][v2]amix=inputs=3:duration=longest:normalize=0[v];[g][m][v]amix=inputs=3:duration=first:normalize=0[a]" \
+$NARR_FX;[v]asplit=2[v1][vk];[g][m]amix=inputs=2:duration=first:normalize=0[bed];\
+[bed][vk]sidechaincompress=threshold=0.02:ratio=6:attack=40:release=500[ducked];[ducked][v1]amix=inputs=2:duration=first:normalize=0[a]" \
 	-map 0:v -map "[a]" -vf "scale=in_range=pc:out_range=tv,format=yuv420p" -color_range tv \
 	-colorspace bt709 -color_primaries bt709 -color_trc bt709 \
 	-c:v libx264 -profile:v high -level 4.0 -preset slow -crf 23 -maxrate 3500k -bufsize 7000k -tag:v avc1 \

tools/trailer/make_trailer_voice.py

@@ -1,7 +1,10 @@
-"""Old Tam's lines for the trailer (tools/trailer/make_trailer.sh mixes them
-in): short ones, in his voice (the same clone as tools/voice/make_voice.py),
-each checked by Whisper and the best of three takes kept. Output:
-tools/trailer/voice/<name>.ogg. Run with the voice environment:
+"""Old Tam narrates the trailer (tools/trailer/make_trailer.sh mixes it in):
+one line for each shot, in his voice (the same clone as
+tools/voice/make_voice.py), each checked by Whisper, the best of three takes
+kept, and sped up a touch if it would run past the start of the next line.
+Output: tools/trailer/voice/<NN>.ogg and lines.json ([{file, shot, at, text}]:
+`at` is seconds into its shot; make_trailer.sh reads the shots' start times
+from the render log). Run with the voice environment:
   tools/voice/.venv-qwen/bin/python tools/trailer/make_trailer_voice.py
 """
 import json, os, subprocess, sys, tempfile
@@ -11,12 +14,20 @@ sys.path.insert(0, os.path.join(HERE, "..", "voice"))
 import make_voice as mv   # (its voice, its checks)
 
 OUT = os.path.join(HERE, "voice")
-## name -> what he says (make_trailer.sh times each against the shots).
-LINES = {
-    "dawn": "Ha! Awake at last! Washed up on my island, eh?",
-    "wolves": "Hush now. See it, there, at the edge of the light?",
-    "cove": "Well, you made it down!",
-}
+## The shots (trailer.gd), their lengths in seconds, and what he says over
+## each, starting `at` seconds in.
+SHOTS = [7.5, 5.0, 6.0, 5.5, 4.5, 5.5, 6.0, 6.0]
+LINES = [
+    (0, 2.0, "Welcome to Cinderwild. You wake on a wild island, washed up with nothing but the clothes on your back."),
+    (1, 0.3, "Gather what you can. Hunt, forage, and cook over an open fire."),
+    (2, 0.3, "Build yourself a home, plank by plank, before the nights draw in."),
+    (3, 0.3, "Because when the sun goes down, the wolves come out. Keep your fire burning."),
+    (4, 0.3, "Weather the storms, and the hard winters."),
+    (5, 0.3, "Then dig down, through the sea cliffs."),
+    (6, 0.3, "And find out what lies beyond."),
+    (7, 0.6, "Play alone, or with your friends, on your own server. Cinderwild."),
+]
+GAP = 0.25   # seconds of quiet before the next line
 
 
 def main():
@@ -25,26 +36,43 @@ def main():
     import mlx_whisper
     from mlx_audio.tts.utils import load_model
     os.makedirs(OUT, exist_ok=True)
+    for f in os.listdir(OUT):
+        if f.endswith(".ogg"):
+            os.remove(os.path.join(OUT, f))
     model = load_model(mv.MODEL)
     sr = getattr(model, "sample_rate", 24000)
     tmp = tempfile.mkdtemp()
-    for name, text in LINES.items():
+    out = []
+    for n, (shot, at, text) in enumerate(LINES):
+        # Room: to the next line's start (or the end of the film).
+        if n + 1 < len(LINES):
+            nshot, nat, _ = LINES[n + 1]
+            room = sum(SHOTS[shot:nshot]) - at + nat - GAP
+        else:
+            room = sum(SHOTS[shot:]) - at - 0.5
         best = None
         for k in range(3):
             res = list(model.generate(text=text, ref_audio=mv.REF_AUDIO, ref_text=mv.REF_TEXT))
             audio = np.concatenate([np.array(r.audio) for r in res])
-            wav = os.path.join(tmp, "%s_%d.wav" % (name, k))
+            wav = os.path.join(tmp, "%02d_%d.wav" % (n, k))
             sf.write(wav, audio, sr)
             heard = mlx_whisper.transcribe(wav, path_or_hf_repo=mv.WHISPER, language="en")["text"].strip()
-            score = mv.match(text, heard)
+            secs = len(audio) / sr
+            # Words first; then the one that fits.
+            score = mv.match(text, heard) - (0.05 if secs > room * 1.15 else 0.0)
             if best is None or score > best[0]:
-                best = (score, wav, len(audio) / sr)
-            if score >= 0.98:
+                best = (score, wav, secs)
+            if score >= 0.98 and secs <= room:
                 break
-        subprocess.run(["ffmpeg", "-v", "error", "-y", "-i", best[1], "-ac", "2", "-c:a", "vorbis", "-strict", "-2",
-                        "-q:a", "5", os.path.join(OUT, name + ".ogg")], check=True)
-        print("  %s (%.1f s, words %.0f%%)" % (name, best[2], best[0] * 100), flush=True)
-    json.dump(LINES, open(os.path.join(OUT, "lines.json"), "w"), indent=2)
+        score, wav, secs = best
+        tempo = min(max(secs / room, 1.0), 1.2) if room > 0 else 1.0
+        name = "%02d.ogg" % n
+        subprocess.run(["ffmpeg", "-v", "error", "-y", "-i", wav, "-af", "atempo=%.3f" % tempo, "-ac", "2",
+                        "-c:a", "vorbis", "-strict", "-2", "-q:a", "5", os.path.join(OUT, name)], check=True)
+        out.append({"file": name, "shot": shot, "at": at, "text": text})
+        print("  %s shot %d: %.1f s%s (room %.1f s, words %.0f%%)" % (name, shot, secs / tempo,
+              " (x%.2f)" % tempo if tempo > 1.0 else "", room, score * 100), flush=True)
+    json.dump(out, open(os.path.join(OUT, "lines.json"), "w"), indent=2)
 
 
 if __name__ == "__main__":

tools/trailer/voice/lines.json

@@ -1,5 +1,50 @@
-{
-  "dawn": "Ha! Awake at last! Washed up on my island, eh?",
-  "wolves": "Hush now. See it, there, at the edge of the light?",
-  "cove": "Well, you made it down!"
-}
+[
+  {
+    "file": "00.ogg",
+    "shot": 0,
+    "at": 2.0,
+    "text": "Welcome to Cinderwild. You wake on a wild island, washed up with nothing but the clothes on your back."
+  },
+  {
+    "file": "01.ogg",
+    "shot": 1,
+    "at": 0.3,
+    "text": "Gather what you can. Hunt, forage, and cook over an open fire."
+  },
+  {
+    "file": "02.ogg",
+    "shot": 2,
+    "at": 0.3,
+    "text": "Build yourself a home, plank by plank, before the nights draw in."
+  },
+  {
+    "file": "03.ogg",
+    "shot": 3,
+    "at": 0.3,
+    "text": "Because when the sun goes down, the wolves come out. Keep your fire burning."
+  },
+  {
+    "file": "04.ogg",
+    "shot": 4,
+    "at": 0.3,
+    "text": "Weather the storms, and the hard winters."
+  },
+  {
+    "file": "05.ogg",
+    "shot": 5,
+    "at": 0.3,
+    "text": "Then dig down, through the sea cliffs."
+  },
+  {
+    "file": "06.ogg",
+    "shot": 6,
+    "at": 0.3,
+    "text": "And find out what lies beyond."
+  },
+  {
+    "file": "07.ogg",
+    "shot": 7,
+    "at": 0.6,
+    "text": "Play alone, or with your friends, on your own server. Cinderwild."
+  }
+]