Workflow: Game Audio Designer

Persona: Sam — sole audio designer at a 30-person studio shipping a sci-fi shooter. Needs hundreds of SFX variants, adaptive music stems, ambient loops, and VO scratch tracks.

---

What you'll generate

• Weapon SFX variants (20 in bulk)
• Combat music loop (adaptive stem)
• Ambient space station soundscape
• VO scratch track for a villain line
• Voice clone for ADR from a reference recording

---

Prerequisites

pip install "assgen[inference]"
assgen server status   # confirm CUDA device, confirm audio models are available
assgen tasks --domain audio   # see all audio tasks and assigned models

---

Sound effects

Single SFX

assgen gen audio sfx generate \
  --prompt "plasma rifle shot, sci-fi, sharp crack, with energy tail" \
  --duration 2.0 \
  --wait

Output: sfx_<id>.wav — 44.1 kHz stereo WAV.

Batch SFX variants (20 in parallel)

Generating weapon variant sets is the most common audio task. Submit all at once and let the server process the queue:

#!/bin/bash
VARIANTS=(
  "plasma rifle shot, sci-fi, sharp crack"
  "plasma rifle shot, sci-fi, deep resonant boom"
  "plasma rifle shot, sci-fi, high pitched whine"
  "plasma rifle charged shot, sci-fi, building energy release"
  "plasma rifle near-miss, sci-fi, Doppler whoosh"
  "plasma rifle dry fire, sci-fi, click and hiss"
  "plasma rifle overheat, sci-fi, sizzle and steam"
  "plasma rifle reload mechanism, sci-fi, metallic clicks"
  "plasma rifle impact on metal, sci-fi, ricochet"
  "plasma rifle impact on flesh, sci-fi, wet thud"
)

for VARIANT in "${VARIANTS[@]}"; do
  assgen gen audio sfx generate \
    --prompt "$VARIANT" \
    --duration 2.0 &
done
wait
echo "✓ Queued $(assgen jobs list | grep pending | wc -l) jobs"
echo "   Monitor: assgen jobs list"

Duration guidance

• Impact/hit sounds: 0.5–1.0 s
• Gunshots: 1.0–2.0 s
• Ambience stingers: 3.0–5.0 s
• AudioLDM2 works best under 10 seconds; longer clips may lose coherence.

---

Music

Combat loop (single stem)

assgen gen audio music loop \
  --prompt "tense sci-fi combat music, electronic drums, 120bpm, driving bass, minor key, loopable" \
  --duration 30 \
  --wait

Output: music_<id>.wav

Adaptive music stems

For a reactive music system, generate stems that share BPM and key so they blend at runtime:

SHARED="120bpm, D minor, sci-fi electronic, adaptive game music stem"

# Low-intensity: ambient underscore
assgen gen audio music loop \
  --prompt "quiet ambient exploration, no drums, $SHARED" \
  --duration 60 &

# Mid-intensity: tension ramp
assgen gen audio music loop \
  --prompt "building tension, light percussion, $SHARED" \
  --duration 60 &

# High-intensity: full combat
assgen gen audio music loop \
  --prompt "full combat, heavy drums, aggressive bass, $SHARED" \
  --duration 60 &

wait

BPM and key consistency

Always include the same BPM and key string across all stems. MusicGen conditions on these tokens — matching them across prompts is the best way to get stems that blend without pitch or timing drift.

Loopless click check

Open the output in Audacity, select 0.5 s around the loop point (end→start), and listen for a click. If present, trim 10–20 ms from each end with a fade-in/fade-out of 5 ms.

---

Ambient audio

assgen gen audio ambient generate \
  --prompt "deep space station ambience, low mechanical hum, distant metallic resonance, pressurised atmosphere" \
  --duration 60 \
  --wait

Output: ambient_<id>.wav

Ambient loops benefit from longer durations (60–120 s) to avoid audible repetition. Generate 2–3 variants and crossfade between them at runtime.

---

Voice and VO

TTS scratch track

assgen gen audio voice tts \
  "Your species ends today." \
  --voice villain \
  --wait

Output: tts_<id>.wav — Bark-generated speech.

Scratch track quality

Bark is excellent for scratch VO and placeholder audio. It is not suitable for final VO — the prosody can be robotic and the speaker identity is not stable across lines. Use it to block out timing; replace with a voice actor for ship.

Batch NPC dialogue

# dialog.json
# [
#   {"character": "guard", "line": "Halt! Who goes there?"},
#   {"character": "guard", "line": "Move along, citizen."},
#   {"character": "scientist", "line": "The readings are off the charts!"}
# ]

assgen gen audio voice dialog dialog.json \
  --voice npcs.yaml \
  --wait

Voice clone (ADR / scratch replacement)

When a director has recorded a temp track and you need consistent voice identity:

assgen gen audio voice clone \
  --input director_reference.wav \
  --text "We need to retreat now. Fall back to the extraction point." \
  --wait

Output: clone_<id>.wav

Voice clone quality

Clone output is suitable for internal scratch and ADR reference only. It is not production-quality and must not be used as final shipped audio without the voice actor's explicit consent.

---

Organising your audio output

Rename and organise generated files before importing into your audio middleware:

mkdir -p sfx/weapons music/combat music/ambient voice/villain

# Rename to something meaningful
mv sfx_abc12345.wav sfx/weapons/plasma_rifle_shot_v01.wav
mv music_def67890.wav music/combat/combat_full_120bpm.wav

In Wwise or FMOD: import the WAVs, set loop points on music and ambient files, and build your adaptive music container with the stems as blend tracks.

---

Known limitations

Gap	Status	Notes
Longer SFX (> 10 s)	Works with degradation	AudioLDM2 quality drops past 10 s; stitch two clips if needed
Adaptive stem key detection	Manual	Include key in prompt; no auto-detection yet
Spatial audio (ambisonics)	Not in catalog	Convert stereo → ambisonics in Reaper/SPARTA post-generation
Final-quality TTS	Not in scope	Bark is scratch-only; use a professional voice actor for ship

---

Next steps

• CLI Reference — sfx generate, music loop, voice tts, voice clone flags
• Configuration — swap in a larger MusicGen model or a different AudioLDM2 variant for higher quality
• Server Setup — run inference on a dedicated machine, submit from your DAW workstation