evo-video-filler-word-remover

evo-video-filler-word-remover

Purpose

Detects filler words and phrases in interview videos using speech recognition (faster-whisper), outputs timestamped annotations as JSON, and stitches all filler word clips into a single compilation video.

Key Concepts

• Uses faster-whisper (CTranslate2-based Whisper) for CPU-optimized transcription
• small model with int8 quantization: optimal tradeoff between accuracy and CPU speed (~1GB RAM, 1.5-3x real-time)
• Word-level timestamps via DTW cross-attention alignment (~20ms resolution)
• VAD filter (Silero) enabled to prevent hallucination on silence
• condition_on_previous_text=False to prevent the model from skipping disfluencies in favor of grammatically clean output
• Dual-pass detection: multi-word phrases first (greedy longest match), then single-word fillers
• Text normalization: lowercase, strip punctuation, handle Whisper's leading-space tokenization artifacts
• Hesitation equivalence: um, uh, hum, hmm, mhm, umm variants all recognized
• Frame-accurate cutting with re-encoding (libx264/aac) — stream copy cannot achieve sub-second precision due to keyframe spacing
• MPEG-TS intermediate containers for seamless concat demuxer stitching
• Timestamp padding (150ms before, 100ms after) to capture full acoustic envelope of filler utterances
• Overlapping segment merging (gap < 100ms threshold) to avoid micro-cuts

Filler Words Detected

Single-word

um, uh, hum, hmm, mhm, like, yeah, so, well, okay, basically

Multi-word (n-gram scan)

you know, i mean, kind of, i guess

Usage

import sys
sys.path.insert(0, '/app/environment/skills/evo-video-filler-word-remover/scripts')
from utils import run_pipeline

# Full pipeline
fillers = run_pipeline('/root/input.mp4', '/root/annotations.json', '/root/output.mp4')

# Or step by step:
from utils import (extract_audio, transcribe_audio, detect_fillers,
                    compute_segments, extract_and_stitch)

audio_path = extract_audio('/root/input.mp4')
words = transcribe_audio(audio_path, model_size='small')
fillers = detect_fillers(words)
segments = compute_segments(fillers)
extract_and_stitch('/root/input.mp4', segments, '/root/output.mp4')