v3.2.0

Added

• Direct TTS HTTP client — Speaker now uses a direct Faraday-based SpeechClient instead of the non-existent RubyLLM.speak(), making text-to-speech work at runtime
• OpenAI TTS support — Full support for tts-1 and tts-1-hd models via OpenAI's /v1/audio/speech endpoint
• ElevenLabs TTS support — Full support for all ElevenLabs model generations: v1 (eleven_monolingual_v1, eleven_multilingual_v1), v2 (eleven_multilingual_v2, eleven_turbo_v2, eleven_flash_v2), v2.5 (eleven_turbo_v2_5, eleven_flash_v2_5), and v3 (eleven_v3)
• ElevenLabs configuration — New elevenlabs_api_key and elevenlabs_api_base config options for ElevenLabs API access
• TTS pricing cascade — SpeechPricing module with 3-tier pricing resolution: LiteLLM JSON (future-proof), user-configurable tts_model_pricing overrides, and hardcoded per-model fallbacks
• tts_model_pricing config — Override TTS pricing per model via config.tts_model_pricing = { "tts-1" => 0.015 }
• default_tts_cost config — Fallback cost per 1K characters for unknown models
• UnsupportedProviderError — Raised when a TTS provider is not :openai or :elevenlabs
• SpeechApiError — Raised when a TTS API returns an HTTP error, includes status and response_body attributes

Fixed

• Speaker runtime crash — RubyLLM.speak() does not exist in ruby_llm v1.12.0; replaced with direct HTTP client
• ElevenLabs had no backing implementation — ElevenLabs was referenced in the DSL but had no actual API integration

Changed

• Speaker specs rewritten to use WebMock HTTP stubs instead of method-level mocks, testing real code paths
• Removed MockSpeechResponse and RubyLLM.speak mock from test support
• calculate_cost now delegates to SpeechPricing module with differentiated ElevenLabs pricing (flash/turbo at $0.15/1K, premium at $0.30/1K)