AGENT SKILL

text-to-speech

Produce clean, aligned, delivery-ready TTS segments with loudness normalization, boundary cleanup, and export specs for dubbing.

License 许可
Apache-2.0
最近更新
2026-11-14
Stars
★ 0
Forks
1

01 前言 Front Matter

name
text-to-speech
description
Produce clean, aligned, delivery-ready TTS segments with loudness normalization, boundary cleanup, and export specs for dubbing.

03 SKILL.md 正文

Steps

  1. Generate TTS audio per segment using an appropriate engine (neural offline or cloud), then confirm the TTS output’s native sample rate.
  2. Clean up each segment (lightweight, consistent chain):
    • High-pass around ~20 Hz (rumble/DC removal)
    • Optional low-pass around ~16 kHz (reduce digital fizz)
    • Add short fades at boundaries (e.g., ~50 ms) to prevent clicks
  3. Control duration to fit the timing window:
    • If shorter than the window: pad silence
    • If longer: apply gentle rate_adjust (small tempo change) or trim carefully
  4. Loudness workflow (ITU-R BS.1770-4 target behavior):
    • Measure loudness with ebur128 for reporting/validation.
    • Normalize as a final step with loudnorm (e.g., I=-23, TP=-1.5, LRA=11).
    • Do not rely on loudnorm options that may not exist in the environment (notably measure_only=1 on some FFmpeg builds); use ebur128 for measurement and loudnorm for normalization.
  5. Export delivery audio specs:
    • Ensure output is 48,000 Hz, mono, and a suitable PCM WAV for segment deliverables (e.g., pcm_s16le).

Expected Result

Each TTS segment is clean, click-free, window-aligned with controlled duration (rate_adjust/pad_silence/trim), normalized to a consistent loudness target, and exported as 48 kHz mono audio suitable for muxing into the final dubbed MP4.

04 同仓库更多 Skills