Data Generation & Augmentation¶

How WakeBuilder generates diverse training data from minimal recordings.

Overview¶

WakeBuilder transforms a few voice recordings into thousands of training samples:

flowchart TB
    subgraph "User Input"
        A[1-5 Recordings]
    end

    subgraph "Augmentation Pipeline"
        B[TTS Synthesis<br/>440+ voices]
        C[Speed Variations<br/>4 speeds]
        D[Pitch Shifting<br/>5 pitches]
        E[Volume Changes<br/>4 levels]
        F[Noise Addition<br/>Various SNRs]
    end

    subgraph "Output"
        G[5,000+<br/>Training Samples]
    end

    A --> B
    B --> C --> D --> E --> F --> G

TTS Synthesis¶

Sample Multiplication¶

Each wake word is synthesized across multiple TTS engines:

Engine	Voices	× Speeds	× Volumes	Samples
Piper	~87	3	4	~1,044
Edge	~302	2	4	~2,416
Kokoro	~40	2	4	~320
Coqui	~10	2	4	~80
Total				~3,860

Voice Diversity¶

TTS provides diversity in:

Gender: Male and female voices
Accent: American, British, Australian, Indian
Age: Various perceived ages
Style: Neutral, expressive, formal

Audio Augmentation¶

Speed Variation¶

Speed	Purpose
0.9x	Slower speakers
1.0x	Normal speed
1.1x	Faster speakers
1.5x	Very fast speech

Pitch Shifting¶

Semitones	Purpose
-2	Lower voices
-1	Slightly lower
0	Original
+1	Slightly higher
+2	Higher voices

Volume Variation¶

Gain (dB)	Purpose
-6 dB	Quiet speech
-3 dB	Moderately quiet
0 dB	Normal
+3 dB	Loud speech

Noise Addition¶

SNR (dB)	Environment
-20 dB	Very noisy
-10 dB	Moderately noisy
-5 dB	Light noise
0 dB	Clean (no noise)

Negative Sample Generation¶

Categories¶

Type	Source	Purpose
Hard negatives	TTS	Similar-sounding words
Real negatives	Cache	Speech, music, ambient
Noise	Generator	Pure noise, silence

Hard Negatives (Critical)¶

Phonetically similar words prevent false positives:

Wake word: "jarvis"

Hard negatives generated:
- Pure prefixes: ja, jar, jarv, jarvi
- Extensions: jarvey, jarvice, jarvson
- Suffixes: arvis, rvis, vis
- Edit distance 1: javis, jarvs, jervis
- Phonetic: gervice, chervice

Real Negatives (from UNAC)¶

Pre-processed chunks from diverse audio:

Category	Description	Chunks
Music	Various genres	~15,000
Speech	Random speech	~20,000
Ambient	Environmental	~8,000
Silence	Near-silent	~4,000
Total		~47,000

Sample Balancing¶

Typical Distribution¶

With default settings (5,000 positives):

Type	Count	Ratio
Positives (TTS + augmented)	5,000	1.0x
Real negatives	10,000	2.0x
Hard negatives	20,000	4.0x
Total	35,000

Train/Validation Split¶

Set	Percentage	Purpose
Training	75%	Model learning
Validation	25%	Performance evaluation

Important: Validation uses unseen TTS voices to test generalization.

Configuration¶

Data Generation Settings¶

Setting	Default	Range
`target_positive_samples`	5,000	3,000 - 8,000
`negative_ratio`	2.0	1.0 - 3.0
`hard_negative_ratio`	4.0	2.0 - 6.0
`use_tts_positives`	true	-
`use_real_negatives`	true	-
`use_hard_negatives`	true	-

Augmentation Settings¶

Setting	Default	Purpose
`speed_range`	[0.9, 1.1]	Normal variations
`speed_fast`	[1.3, 1.5]	Fast speech
`pitch_range`	[-2, +2]	Semitones
`volume_range`	[0.7, 1.3]	Gain multiplier
`noise_snr_range`	[-20, -5]	dB

Processing Flow¶

Training Data Preparation¶

sequenceDiagram
    participant User
    participant Trainer
    participant TTS
    participant Augmenter
    participant Cache

    User->>Trainer: Recordings + Wake Word

    par Positive Generation
        Trainer->>TTS: Synthesize wake word
        TTS-->>Trainer: 440+ voice samples
        Trainer->>Augmenter: Augment all samples
        Augmenter-->>Trainer: Speed, pitch, volume variations
    end

    par Negative Generation
        Trainer->>TTS: Hard negatives (similar words)
        TTS-->>Trainer: Phonetically similar audio
        Trainer->>Cache: Load real negatives
        Cache-->>Trainer: Pre-processed chunks
    end

    Trainer->>Trainer: Balance dataset
    Trainer->>Trainer: Split train/validation
    Trainer-->>User: Ready for training

Reference¶

Source files:

src/wakebuilder/audio/augmentation.py - DataAugmenter class
src/wakebuilder/audio/negative_generator.py - Hard negative generation
src/wakebuilder/audio/real_data_loader.py - Cache loading
src/wakebuilder/tts/ - TTS providers