TTS Research: Alternatives and Commercial Licensing Analysis for Gromozeka

[Lewik]

TTS Research: Alternatives and Commercial Licensing Analysis for Gromozeka

Research Context

Investigation of TTS options for Gromozeka desktop application focusing on:

• Commercial licensing compatibility (for potential product sales)
• Russian language support quality
• Performance and latency requirements
• Integration complexity with Kotlin/JVM
• Cost analysis for desktop chat application usage

Current Implementation Analysis

Existing TTS Setup (Spring AI + OpenAI)

Location: bot/src/jvmMain/kotlin/com/gromozeka/bot/services/TtsService.kt

Current Configuration:

• Model: gpt-4o-mini-tts (cost-optimized choice)
• Cost: /bin/zsh.60/1M characters (25x cheaper than tts-1)
• Voice: OpenAI Alloy voice, MP3 format, 1.0x speed
• Integration: Well-integrated with chat streaming, TTS queue management, cancellation support
• Playback: macOS afplay command
• Monthly cost: < USD for typical usage patterns

Architecture Strengths:

• Proper queue management with cancellation (TTSQueueService.kt)
• Integration with streaming chat responses
• JSON-structured responses from Claude for TTS text extraction
• Already working and stable

Commercial Licensing Analysis

NOT Commercial Friendly:

1. macOS native TTS (say command)

   • Apple''s license restricts System Voices to personal, non-commercial use only
   • Excellent Russian voice Milena available
   • Simple integration: say -v Milena текст
   • Use case: Development/prototyping only

2. Piper TTS

   • License changed: MIT → GPL-3.0 (viral license)
   • Repository moved: rhasspy/piper → github.com/OHF-Voice/piper1-gpl
   • Requires entire application to be open source if used
   • Performance: 20-30ms generation (excellent)
   • Quality: Google TTS level, superior Russian support

3. eSpeak-NG

   • GPL v3+ license (viral)
   • Russian phonemes available but robotic quality
   • brew install espeak-ng

Commercial Friendly:

1. Festival TTS (Only viable open-source option)

   • License: BSD-like (allows commercial use)
   • Cost: Free
   • Quality: Basic/robotic sound
   • Russian support: festvox-ru package available
   • Installation: brew install festival + Russian voice package from festvox.org
   • Use case: Offline fallback option

2. Google Cloud TTS (Recommended cloud option)

   • Free tier: 1M WaveNet characters/month + 4M Standard characters/month
   • Pricing after free tier: /1M WaveNet, /1M Standard
   • Quality: Excellent Russian neural voices (multiple options)
   • Latency: Moderate (API calls)
   • Integration: Java SDK with Kotlin compatibility

3. Azure TTS

   • Free tier: 500K characters/month
   • Pricing: Similar to Google (~/1M neural voices)
   • Russian voices: SvetlanaNeural and others
   • Quality: Very good

4. Amazon Polly

   • Pricing: .20/1M standard, .80/1M neural
   • Russian support: Available
   • Quality: Good

Performance Comparison

Latency Analysis:

• Piper TTS: 20-30ms (100x faster, but GPL licensing issue)
• Current OpenAI: 1-3 seconds + network latency
• Cloud TTS services: 1-2 seconds typical
• macOS native: Very fast (~200ms, but non-commercial)
• Festival TTS: Fast local processing (~500ms)

Russian Language Quality Ranking:

1. Google Cloud TTS: Best neural voices for Russian
2. macOS Milena voice: Excellent but non-commercial
3. Piper TTS: Very good Russian support (but GPL)
4. Azure TTS: Good Russian neural voices
5. Current OpenAI: Decent, optimized for English
6. Amazon Polly: Adequate Russian support
7. Festival TTS: Basic quality, robotic sound
8. eSpeak-NG: Robotic, basic phonemes

Implementation Recommendations

Phase 1: Current State (Keep as-is)

• Rationale: OpenAI gpt-4o-mini-tts is already optimal for cost/quality balance
• Action: No changes needed, current implementation is excellent

Phase 2: Quality Enhancement (Optional)

• Primary: Add Google Cloud TTS integration for better Russian voices
• Implementation: Use 1M free characters/month, fallback to OpenAI
• Benefits: Superior Russian voice quality + cost optimization

Phase 3: Offline Capability (Future)

• Add: Festival TTS as offline fallback option
• Trade-off: Basic quality but no internet dependency
• Use case: Privacy-focused users or network issues

Key Insights

1. Current choice validated: OpenAI gpt-4o-mini-tts is actually the most cost-effective option in 2025
2. GPL license trap: Many promising open-source TTS engines became GPL, making them unsuitable for commercial products
3. Google Cloud opportunity: 1M free characters/month provides significant value for experimentation
4. Hybrid approach optimal: Combine cloud quality with offline fallback options

Action Items for Implementation

• Research Google Cloud TTS Java SDK integration with Spring AI
• Create TTS provider abstraction interface
• Implement Google Cloud TTS provider with free tier management
• Add configuration for TTS engine selection
• Test Festival TTS installation and Russian voice quality
• Document licensing considerations for each TTS option

---

🤖 Generated with Claude Code