TTS Providers¶

WakeBuilder uses multiple Text-to-Speech engines to generate diverse training data.

Overview¶

TTS (Text-to-Speech) synthesis is critical for data augmentation. By generating synthetic voice samples using multiple TTS engines, WakeBuilder creates a diverse training dataset from just a few user recordings.

graph TB
    subgraph "Wake Word Text"
        WW[/"jarvis"/]
    end

    subgraph "TTS Providers"
        PIPER[Piper TTS<br/>~87 voices]
        EDGE[Edge TTS<br/>~302 voices]
        KOKORO[Kokoro TTS<br/>~40+ voices]
        COQUI[Coqui TTS<br/>~10 voices]
    end

    subgraph "Synthesis"
        SPD[Speed Variations<br/>0.9x, 1.0x, 1.1x, 1.5x]
        VOL[Volume Variations<br/>-6dB to +3dB]
    end

    subgraph "Output"
        SAMPLES[Thousands of<br/>Audio Samples]
    end

    WW --> PIPER --> SPD
    WW --> EDGE --> SPD
    WW --> KOKORO --> SPD
    WW --> COQUI --> SPD
    SPD --> VOL --> SAMPLES

Piper TTS¶

Overview¶

Piper is a fast, high-quality offline neural TTS system. It's the primary TTS engine for WakeBuilder due to its quality and local operation.

Characteristics¶

Voice Diversity¶

Piper voices include:

• Genders: Male and female
• Accents: American, British, Australian, Indian
• Ages: Various (young, middle-aged, elderly)
• Styles: Neutral, expressive, formal

Source Reference¶

• Module: src/wakebuilder/tts/generator.py
• Class: TTSGenerator
• Voice Directory: tts_voices/
• Voice List: scripts/piper_tts_voices.json

Usage Example¶

from wakebuilder.tts import TTSGenerator

generator = TTSGenerator()
audio, sample_rate = generator.synthesize(
    text="jarvis",
    voice_name="en_US-amy",
    length_scale=1.0  # Speed: <1.0 faster, >1.0 slower
)

Edge TTS¶

Overview¶

Edge TTS provides access to Microsoft's neural TTS voices used in Edge browser. It offers exceptional quality and a large voice selection.

Characteristics¶

Voice Categories¶

Source Reference¶

• Module: src/wakebuilder/tts/edge_generator.py
• Class: EdgeTTSGenerator
• Voice List: scripts/list of voices available in Edge TTS.txt

Usage Example¶

from wakebuilder.tts.edge_generator import EdgeTTSGenerator

generator = EdgeTTSGenerator()
audio, sample_rate = generator.synthesize(
    text="jarvis",
    voice="en-US-JennyNeural",
    speed=1.0  # Speed multiplier
)

Speed Variations¶

Edge TTS uses rate strings:

Kokoro TTS¶

Overview¶

Kokoro is a high-quality neural TTS model with 40+ English voices. It provides excellent quality with both American and British accents.

Characteristics¶

Voice Categories¶

Voice ID Examples¶

Source Reference¶

• Module: src/wakebuilder/tts/kokoro_generator.py
• Class: KokoroTTSGenerator
• Constants: KOKORO_VOICES dictionary

Usage Example¶

from wakebuilder.tts.kokoro_generator import KokoroTTSGenerator

generator = KokoroTTSGenerator(use_gpu=True)
audio, sample_rate = generator.synthesize(
    text="jarvis",
    voice_id="af_heart",
    speed=1.0
)

Coqui TTS¶

Overview¶

Coqui TTS provides high-quality VITS-based synthesis with support for multiple speakers.

Characteristics¶

Available Models¶

Source Reference¶

• Module: src/wakebuilder/tts/coqui_generator.py
• Class: CoquiTTSGenerator
• Model Config: COQUI_MODELS dictionary

Usage Example¶

from wakebuilder.tts.coqui_generator import CoquiTTSGenerator

generator = CoquiTTSGenerator(use_gpu=True)
audio, sample_rate = generator.synthesize(
    text="jarvis",
    model_key="vctk",
    speaker="p225"  # Speaker ID
)

Unified TTS Interface¶

Combined Generation¶

The training pipeline uses all available TTS providers:

# In data augmentation
tts_samples = []

# Piper TTS (primary)
piper_gen = TTSGenerator()
for voice in piper_gen.available_voices:
    audio = piper_gen.synthesize(wake_word, voice)
    tts_samples.append(audio)

# Edge TTS (secondary)
edge_gen = EdgeTTSGenerator()
for voice in edge_gen.voice_ids:
    audio = edge_gen.synthesize(wake_word, voice)
    tts_samples.append(audio)

# Kokoro TTS (tertiary)
kokoro_gen = KokoroTTSGenerator()
for voice in kokoro_gen.voice_ids:
    audio = kokoro_gen.synthesize(wake_word, voice)
    tts_samples.append(audio)

Sample Generation Flow¶

flowchart TB
    WW[Wake Word] --> |"jarvis"| TTS_ALL

    subgraph TTS_ALL["All TTS Providers"]
        P[Piper: ~87 voices]
        E[Edge: ~302 voices]
        K[Kokoro: ~40+ voices]
        C[Coqui: ~10 voices]
    end

    TTS_ALL --> SPEED["Speed Variations<br/>×3-4 each"]
    SPEED --> VOLUME["Volume Variations<br/>×4 each"]
    VOLUME --> TOTAL["Thousands of samples"]

Configuration¶

TTS Settings in config.py¶

# Speed variations applied to TTS output
SPEED_VARIATIONS = [0.9, 1.0, 1.1, 1.5]

# Volume variations in dB
VOLUME_VARIATIONS = [-6, -3, 0, 3]

# Target sample rate (resample TTS output)
TARGET_SAMPLE_RATE = 16000

Enabling/Disabling Providers¶

In the web interface, use the training options:

• Generate TTS Positives: Enable/disable TTS synthesis
• TTS engines are used based on availability

Downloading Voice Models¶

Piper Voices¶

Use the download script:

python scripts/download_voices.py

This downloads:

• All English Piper voice models
• Configuration files
• Total: ~5GB

Other TTS Dependencies¶

Edge, Kokoro, and Coqui models are downloaded on first use.

Comparison Table¶

Troubleshooting¶