Long-Form-Audio-Pipeline

A preprocessing pipeline for preparing long-form audio recordings for speech-to-text (ASR) APIs. This tool optimizes audio files for transcription by reducing file size, normalizing audio levels, and cleaning up the signal before upload.

Why Preprocess?

While many ASR APIs perform server-side preprocessing, this pipeline:

• Reduces upload time and bandwidth by compressing audio before transmission
• Ensures consistent input quality regardless of original recording conditions
• Removes unnecessary data (silence, stereo channels) that don't aid transcription
• Optimizes for speech recognition with filters tuned for voice clarity

Pipeline Stages

The preprocessing script applies the following transformations in order:

Stage	Transformation	Details
1	Silence Detection & Trimming	Detects silence at start/end (threshold: -50dB, min duration: 0.5s) and trims it
2	Stereo to Mono Conversion	Mixes stereo channels to mono (0.5×L + 0.5×R) — ASR doesn't benefit from stereo
3	Sample Rate Conversion	Resamples to 16kHz — the standard rate for most STT systems
4	Loudness Normalization	Applies EBU R128 normalization (I=-16 LUFS, TP=-1.5 dB, LRA=11)
5	High-Pass Filter	Removes low-frequency rumble below 80Hz common in voice recordings
6	Speech Clarity EQ	4-band EQ optimized for voice intelligibility (see below)
7	Dynamic Range Compression	Gentle compression (threshold: -20dB, ratio: 3:1) to even out volume levels
8	Codec Transformation	Encodes to OGG Opus at 48kbps VBR with voip application mode

Speech Clarity EQ Details

The EQ stage applies frequency adjustments specifically tuned for speech intelligibility:

Frequency	Adjustment	Purpose
300 Hz	-3 dB cut	Reduces muddiness in the low-mid range
1.5 kHz	+3 dB boost	Enhances vowel presence and clarity
3 kHz	+4 dB boost	Boosts consonant clarity (the "intelligibility band")
4 kHz+	+2 dB shelf	Adds air and crispness to sibilants

These adjustments were derived from frequency analysis of real voice recordings, targeting the bands that most impact ASR accuracy.

Example Results

Processing a real-world voice recording through this pipeline:

Metric	Before	After
File Size	482.87 MB	16.80 MB
Format	WAV (PCM 16-bit)	OGG (Opus)
Channels	Stereo	Mono
Sample Rate	44100 Hz	16000 Hz
Duration	2870.32s (~47 min)	2869.71s
Compression Ratio	—	28.7×

The 0.6 seconds trimmed from the start was detected silence before speech began.

GUI Application

A desktop GUI is available for easy audio processing and EQ calibration.

Features:

• Drag & drop audio file input
• Automatic output naming (_processed.ogg suffix)
• EQ calibration workflow with sample extraction
• Import/export EQ presets (.conf, .json, .ffmpeg formats)
• View and edit current EQ bands

Tab	Description
Process Audio	Drop files for processing with current EQ preset
EQ Calibration	Extract samples, edit in Audacity, compute custom EQ
Current Preset	View, import, and export EQ presets

Install GUI (Debian/Ubuntu)

sudo apt install ./audio-processor_1.0.0_all.deb

Or run directly:

./install_gui.sh   # Sets up Python venv and PyQt6
./run_gui.sh       # Launch the GUI

Installation

Dependencies

• ffmpeg with libopus support
• ffprobe (included with ffmpeg)
• Standard Unix tools: awk, grep, sed

On Ubuntu/Debian:

sudo apt install ffmpeg

Usage

Quick Start (Single File)

# Process a single file with default EQ
./preprocess_for_stt.sh recording.wav

Batch Processing with Custom EQ

This pipeline supports creating your own EQ preset by editing a sample in your preferred audio editor:

# 1. Extract a 15-second sample for calibration
./extract_sample.sh calibration.wav

# 2. Edit samples/sample_before.wav in your audio editor
#    Apply your preferred EQ adjustments
#    Export as samples/sample_after.wav

# 3. Compute your custom EQ preset
./compute_eq.sh

# 4. Drop files in the 'in/' folder and batch process
./process_batch.sh

Directory Structure

├── in/                    # Drop input files here
├── out/                   # Processed files appear here
├── samples/
│   ├── sample_before.wav  # Raw sample for editing
│   └── sample_after.wav   # Your edited sample (you create this)
├── eq_preset.conf         # Generated EQ preset
├── extract_sample.sh      # Extract calibration sample
├── compute_eq.sh          # Derive EQ from before/after
├── process_batch.sh       # Batch process with custom EQ
└── preprocess_for_stt.sh  # Single-file processor (default EQ)

Supported Input Formats

Any format ffmpeg can decode:

• WAV, FLAC, AIFF (uncompressed)
• MP3, AAC, OGG, M4A (compressed)
• Most other audio formats

Configuration

Edit the variables at the top of preprocess_for_stt.sh:

SAMPLE_RATE=16000           # Target sample rate (Hz)
BITRATE="48k"               # OGG Opus bitrate
SILENCE_THRESHOLD="-50dB"   # Silence detection threshold
SILENCE_DURATION="0.5"      # Minimum silence duration (seconds)

When to Use This

This pipeline is specifically intended for:

• Voice memos and dictation recordings
• Interview recordings
• Podcast audio for transcription
• Meeting recordings
• Any long-form speech content destined for ASR

It's optimized for speech, not music. The high-pass filter and compression settings are tuned for voice clarity.

License

MIT