TTS & SSML Prosody Control - 6-Minute Manim Video

This project creates a comprehensive 5-6 -minute video presentation about "Improving French Synthetic Speech Quality via SSML Prosody Control" using Manim Community v0.18+.

📁 Project Structure

tts_ssml_manim_video/
├── manim.py                 # Main Manim script with all scenes
├── assets/                 # Visual assets extracted from PPT
│   ├── slide_20_img_8.png
│   ├── slide_23_img_7.png
│   └── slide_23_img_8.png
├── extracted_data/         # Data extraction from source files
│   └── data_extraction.json
├── citations.jsonl         # Complete source tracking
└──README.md              # This file

📊 Source Files

• PDF: ICNLSP 2025_P25-1088_camera_ready.pdf - Research paper
• PPT: Text_To_Speech_copy (1).pptx - Course slides

🎬 Video Structure

Total Duration: 5-6 Minutes

Scene Breakdown

Scene	Duration	Description	Source
SceneIntro	30s	Title, authors, paper reference	PDF p.1
SceneBasics	90s	Waveform, spectrogram, pitch/F0	PPT slides 9,12,13,16,22
SceneProblem	75s	TTS expressivity problem	PDF p.1-2
ScenePipeline	90s	Text→SSML→TTS pipeline	PDF p.3 + PPT slide 26
SceneStage1	60s	Break insertion (QwenA)	PDF p.4,7 Table 4
SceneStage2	60s	Prosody prediction (QwenB)	PDF p.4-5
SceneEvalObj	105s	F1, MAE/RMSE metrics	PDF p.6-7 Tables 4-5
SceneEvalSubj	60s	MOS scores, AB test	PDF p.1,6
SceneOutro	30s	Conclusions & future work	PDF p.8
Total	600s	Including transitions

🚀 Execution

Requirements

# Install Manim Community v0.18+
pip install manim

# Verify installation
manim --version

Generate the Full Video

manim -pqh manim.py Main -o video.mp4 --format=mp4 --fps 30 --resolution 1920,1080

Generate Individual Scenes

# Scene 0: Introduction
manim -pqh manim SceneIntro -o intro.mp4 --format=mp4 --fps 30 --resolution 1920,1080
        
# Scene 1: Audio Basics
manim -pqh manim.py SceneBasics -o basics.mp4 --format=mp4 --fps 30 --resolution 1920,1080

# Scene 2: TTS Problem
manim -pqh manim.py SceneProblem -o problem.mp4 --format=mp4 --fps 30 --resolution 1920,1080

# Scene 3: Pipeline
manim -pqh manim.py ScenePipeline -o pipeline.mp4 --format=mp4 --fps 30 --resolution 1920,1080

# Scene 4: Stage 1
manim -pqh manim.py SceneStage1 -o stage1.mp4 --format=mp4 --fps 30 --resolution 1920,1080

# Scene 5: Stage 2
manim -pqh manim.py SceneStage2 -o stage2.mp4 --format=mp4 --fps 30 --resolution 1920,1080

# Scene 6: Objective Evaluation
manim -pqh manim.py SceneEvalObj -o eval_obj.mp4 --format=mp4 --fps 30 --resolution 1920,1080

# Scene 7: Subjective Evaluation
manim -pqh manim.py SceneEvalSubj -o eval_subj.mp4 --format=mp4 --fps 30 --resolution 1920,1080

# Scene 8: Conclusions
manim -pqh manim.py SceneOutro -o outro.mp4 --format=mp4 --fps 30 --resolution 1920,1080

Command Options

• -pqh: Preview, Quality High
• -o video.mp4: Output filename
• --format=mp4: Video format (H.264)
• --fps 30: Frame rate
• --resolution 1920,1080: Full HD resolution

🎨 Visual Design

Theme Colors

• Background: #004178 (dark blue-black)
• Accent Red: #FF0049 (titles, highlights)
• Accent Red: #FF0049 (emphasis, numbers)
• Text: White

Typography

• Clean sans-serif Text() objects
• Font sizes: 16-52pt depending on hierarchy
• Bold weights for emphasis
• Italic for citations

📈 Key Data Points

All data is sourced from the provided files with zero hallucination:

From PDF (ICNLSP 2025)

• Corpus: 14h French, 14 speakers (42% female), 122,303 words
• F₁ Score: 99.24% (QwenA break prediction)
• MAE: Pitch 0.97%, Volume 1.09%, Rate 1.10%
• Break MAE: 132.89 ms
• MOS: 3.20 → 3.87 (p < 0.005)
• Preference: 15 of 18 participants

From PPT (Course Slides)

• Waveform: Time (s) vs. Amplitude (normalized)
• Spectrogram: 20-30ms windows, ~10ms hop, Hann window
• Pitch: Related to F₀ (fundamental frequency)

📝 Citation Summary

Scene-wise Source Distribution

• SceneIntro: PDF page 1
• SceneBasics: PPT slides 9, 12, 13, 16, 22
• SceneProblem: PDF pages 1-2
• ScenePipeline: PDF page 3 + PPT slide 26
• SceneStage1: PDF page 4, 7 (Table 4), Appendix A
• SceneStage2: PDF pages 4-5
• SceneEvalObj: PDF pages 6-7 (Tables 4-5)
• SceneEvalSubj: PDF pages 1, 6 (Section 5.1)
• SceneOutro: PDF page 8 (Sections 6-7)

📚 References

• Paper: Improving French Synthetic Speech Quality via SSML Prosody Control (Ouali et al., ICNLSP 2025)
• Code Repository: https://github.com/hi-paris/Prosody-Control-French-TTS

👥 Authors

• Nassima Ould Ouali

📄 License

This project is licensed under the MIT License.

---

Manim Version: Community v0.18+ Duration: 400s ± 15s Resolution: 1920x1080 @ 30fps Format: H.264 MP4