Drama Audio Soundstage

Drama Audio Soundstage (Sound Engineering & Intelligent Soundtrack Stage)

Stage Position: Short drama sound design and commercial soundtrack stage. No Generic Fluff: We do not explain what an audio track is. We only document audio gating cleanup parameters, language drift interception thresholds, and sidechain ducking mixing curves.

---

1. Runtime Contract & Input/Output

• Input:
  1. Video segments with native audio from Stage 3: segments/{seg_id}.mp4;
  2. High-quality standalone film score BGM: assets/bgm/*.wav.
• Output:
  1. Cleaned segments with bed noise removed: cleaned_segments/{seg_id}.mp4;
  2. Final video with 90s suspense/commercial soundtrack and intelligent sidechain ducking: [Episode Name]_final_ducked.mp4;
  3. ASR language & dialogue accuracy quality checklist.

---

2. Hardcore SOP: Spectral Subtraction Audio Gating

Problem: H3-generated audio often contains persistent low-frequency hums, false wind noise, or harsh environmental bed noise that ruins immersion. Goal: Completely remove continuous bed noise while preserving 100% of dialogue and short action sound effects (sword-draw, table-slap, footsteps).

2.1 Gating Core Parameters & Physical Meaning

When calling scripts/audio_gate.py, use the golden parameter set:

• --bed-q 15: Time-domain sliding window percentile (set to 15th percentile) for precise identification of continuous background noise;
• --alpha-bed 6.0: Deduction strength for continuous bed noise;
• --alpha-fx 1.0: Protection factor for transient sound effects/dialogue; smaller values preserve dialogue better;
• --fx-snr 4.0: SNR threshold for transient detection;
• --quiet-gain 0.03: Attenuation factor for pure silent frames.

Command:

python scripts/audio_gate.py --in segments/seg01.mp4 --out cleaned_segments/seg01.mp4

---

3. Hardcore SOP: ASR Mandarin Language Drift Automated Interception

Problem: Character pronunciation drifts into pseudo-English, garbled code, or non-Chinese noise, severely affecting delivery quality.

3.1 Interception & Criteria

After writing to disk, run a quick Whisper language validation:

1. Extract the first 5 seconds of the audio track for Language ID detection;
2. Pass Condition: language == "zh" and confidence prob >= 0.85;
3. Interception Action: If detected as en or any other language, automatically throw LanguageDriftError and mark the segment as failed, triggering a re-run with a different seed.

---

4. Hardcore SOP: Cinematic Sidechain Dynamic Ducking (Sidechain Ducking)

Anti-Pattern (Negative Example) : Never blindly lay BGM at a fixed volume under the video! The dialogue will be drowned by background music, resulting in zero commercial-grade layering.

4.1 FFmpeg Native sidechaincompress Golden Curve

Use the dialogue track as the sidechain control signal to dynamically suppress the BGM in real time:

• BGM Base Level: bgm_volume = 0.35;
• Threshold: ducking_threshold = 0.06 (action triggers ducking the moment the human voice appears);
• Ducking Depth (Ratio) : ducking_ratio = 5.0 (when dialogue speaks, BGM instantly ducks approximately -14dB, leaving a completely clear frequency band for vocals);
• Attack Time: ducking_attack_ms = 80ms (smooth ducking at the first word of dialogue, no harsh pops);
• Release Time: ducking_release_ms = 400ms (after dialogue ends, the BGM smoothly recovers within 0.4 seconds, heightening tension).

Command:

python scripts/cinematic_audio_mixer.py --input concat_raw.mp4 --output final_ducked.mp4 --bgm assets/bgm/suspense.wav