ADR-038: FFmpeg-First Audio Manipulation

• Status: Accepted
• Date: 2026-01-11
• Authors: Podcast Scraper Team
• Related RFCs: RFC-040

Context & Problem Statement

The pipeline needs to perform complex audio operations: format conversion, resampling, VAD (silence removal), and loudness normalization. Pure Python libraries are often slow or require complex chains of C-extensions.

Decision

We standardize on FFmpeg-First Audio Manipulation.

• The pipeline calls the system ffmpeg binary directly via subprocess.
• We use complex filter chains (e.g., silenceremove, loudnorm) to perform multiple operations in a single pass.

Rationale

• Performance: ffmpeg is highly optimized and far faster than equivalent Python libraries like pydub or librosa.
• Completeness: Supports every codec and filter we might ever need without adding dozens of Python dependencies.
• Reliability: Industry-standard tool with predictable behavior across platforms.

Alternatives Considered

1. Python Libraries (pydub/webrtcvad): Rejected due to performance bottlenecks and fragmented feature sets.

Consequences

• Positive: Extremely fast audio pipelines; simplified Python dependency list.
• Negative: Requires the user to have ffmpeg installed on their system.

References

• RFC-040: Audio Preprocessing Pipeline for Podcast Ingestion