Skip to content

VOICE_GUIDE.md

Latest commit

 

History

History
269 lines (195 loc) · 7.32 KB

VOICE_GUIDE.md

File metadata and controls

269 lines (195 loc) · 7.32 KB

🎭 Voice Guide

Complete guide to available voices, quality levels, and language support.

Voice Quality Levels

🌟 High Quality

  • Best audio quality with natural pronunciation
  • Larger model sizes (50-200MB per voice)
  • Slower generation (2-5 seconds)
  • Best for production content

⚖️ Medium Quality

  • Balanced performance and quality
  • Moderate model sizes (20-50MB per voice)
  • Good generation speed (1-3 seconds)
  • Recommended for most use cases

⚡ Low Quality

  • Fastest generation (0.5-1.5 seconds)
  • Smallest model sizes (10-30MB per voice)
  • Good for development/testing
  • Acceptable quality for many applications

Available Voices by Language

🇺🇸 US English

High Quality

  • ljspeech_high - Female, clear articulation, professional quality
    • Model: en_US-ljspeech-high
    • Best for: Audiobooks, presentations, professional content

Medium Quality

  • amy_medium - Female, warm and friendly

    • Model: en_US-amy-medium
    • Best for: General purpose, customer service

  • lessac_medium - Female, professional tone

    • Model: en_US-lessac-medium
    • Best for: Business applications, announcements

  • ryan_medium - Male, clear and confident

    • Model: en_US-ryan-medium
    • Best for: Narration, educational content

Low Quality (Fast)

  • amy_low - Female, quick generation

    • Model: en_US-amy-low
    • Best for: Testing, development

  • lessac_low - Female, rapid synthesis

    • Model: en_US-lessac-low
    • Best for: Prototyping, bulk generation

  • ryan_low - Male, fast processing

    • Model: en_US-ryan-low
    • Best for: Development, demos

  • danny_low - Male, casual tone

    • Model: en_US-danny-low
    • Best for: Informal content, games

  • kathleen_low - Female, conversational

    • Model: en_US-kathleen-low
    • Best for: Chatbots, interactive apps

🇬🇧 British English

Medium Quality

  • alan_medium - Male, refined British accent

    • Model: en_GB-alan-medium
    • Best for: Formal content, educational material

  • jenny_medium - Female, pleasant British accent

    • Model: en_GB-jenny_dioco-medium
    • Best for: General purpose, customer service

🇫🇷 French

Medium Quality

  • siwis_medium - Female, native French pronunciation
    • Model: fr_FR-siwis-medium
    • Best for: French content, language learning

🇩🇪 German

Medium Quality

  • thorsten_medium - Male, clear German pronunciation
    • Model: de_DE-thorsten-medium
    • Best for: German content, business applications

🇪🇸 Spanish

Medium Quality

  • davefx_medium - Male, native Spanish pronunciation
    • Model: es_ES-davefx-medium
    • Best for: Spanish content, international applications

🇮🇹 Italian

X-Low Quality (Fast)

  • riccardo_x_low - Male, Italian pronunciation
    • Model: it_IT-riccardo_fasol-x_low
    • Best for: Italian content, quick generation

OpenAI Compatible Aliases

For compatibility with OpenAI TTS API, these aliases are available:

  • alloy → lessac_medium (US Female)
  • echo → danny_low (US Male)
  • fable → alan_medium (GB Male)
  • onyx → ryan_medium (US Male)
  • nova → amy_medium (US Female)
  • shimmer → ljspeech_high (US Female, High Quality)

Voice Selection Guidelines

For Different Use Cases

🎙️ Podcasts & Audiobooks

  • Recommended: ljspeech_high, ryan_medium, alan_medium
  • Quality: High or Medium
  • Reason: Natural flow, clear articulation

💼 Business Applications

  • Recommended: lessac_medium, ryan_medium, alan_medium
  • Quality: Medium
  • Reason: Professional tone, clear communication

🎮 Games & Interactive Apps

  • Recommended: danny_low, kathleen_low, amy_low
  • Quality: Low (for speed)
  • Reason: Fast generation for dynamic content

🤖 Chatbots & Voice Assistants

  • Recommended: amy_medium, lessac_medium, jenny_medium
  • Quality: Medium
  • Reason: Friendly, conversational tone

📚 Educational Content

  • Recommended: ryan_medium, alan_medium, ljspeech_high
  • Quality: Medium to High
  • Reason: Clear pronunciation, good for learning

🌍 Multilingual Applications

  • Available Languages: English, French, German, Spanish, Italian
  • Recommendation: Use native language voices for best pronunciation

Technical Specifications

Audio Format Details

  • Sample Rate: 22,050 Hz (standard)
  • Bit Depth: 16-bit
  • Channels: Mono
  • Output Formats: WAV (native), MP3 (converted)

Model Information

  • Engine: PiperTTS with ONNX runtime
  • Neural Architecture: VITS (Variational Inference Text-to-Speech)
  • Phonemization: Language-specific phoneme mapping
  • Training Data: High-quality speech datasets

Performance Comparison

Voice Quality Generation Time File Size (10s audio) Model Size Use Case
High 2-5 seconds ~350KB (WAV) 50-200MB Production
Medium 1-3 seconds ~350KB (WAV) 20-50MB General use
Low 0.5-1.5 seconds ~350KB (WAV) 10-30MB Development

Note: File sizes for WAV format are consistent as they depend on audio length, not model quality. Quality affects generation time and audio fidelity.

Adding Custom Voices

Requirements

  1. ONNX model file (.onnx)
  2. Configuration file (.onnx.json)
  3. Compatible with PiperTTS

Installation Steps

  1. Add voice files:
# Copy to voices directory
cp custom-voice.onnx voices/
cp custom-voice.onnx.json voices/
  1. Update configuration:
# Add to config/voice_to_speaker.yaml
custom_voice: custom-voice
  1. Restart services:
docker compose restart

Voice Sources

  • Hugging Face: rhasspy/piper-voices
  • Community Models: Check PiperTTS community for additional voices
  • Custom Training: Train your own with PiperTTS training tools

Troubleshooting Voice Issues

Voice Not Found Error

# Check available voices
curl http://localhost:8000/voices

# Verify voice files exist
docker exec tts-server ls -la /app/voices/

# Check configuration
docker exec tts-server cat /app/config/voice_to_speaker.yaml

Poor Audio Quality

  1. Try higher quality version of the same voice
  2. Check model file integrity (re-download if corrupted)
  3. Verify JSON configuration matches ONNX model

Slow Generation

  1. Use lower quality voices for faster generation
  2. Enable caching for repeated text
  3. Use WAV format to avoid conversion overhead

Foreign Language Issues

  1. Use native language voices for best pronunciation
  2. Check text encoding (UTF-8 recommended)
  3. Verify phoneme mapping in model configuration

Voice Licensing

All included voices are based on open-source models:

  • License: Various open-source licenses (check individual model sources)
  • Commercial Use: Generally permitted (verify specific license)
  • Attribution: Check model documentation for attribution requirements

Important: Always verify licensing terms for your specific use case, especially for commercial applications.