#pragma once // The two speech seams of a voice pipeline. Backends implement them: // classic Cloud TTS/STT (TextToSpeech, SpeechToText) or Gemini audio // (GeminiSynthesizer, GeminiRecognizer). Callers never know which. #include namespace molo::google { enum class AudioEncoding { kMp3, kLinear16, kOggOpus }; const char* audio_encoding_name(AudioEncoding encoding); // Google enum name. const char* audio_encoding_mime(AudioEncoding encoding); struct SynthesisRequest { std::string text; bool ssml = false; // Classic Cloud TTS only. std::string language = "nl-NL"; std::string voice; // Empty: backend default. AudioEncoding encoding = AudioEncoding::kMp3; // Classic Cloud TTS only. double speaking_rate = 1.0; // 0.25 .. 4.0, classic Cloud TTS only. }; struct Audio { std::string bytes; std::string mime; }; enum class SpeechEncoding { kWebmOpus, kOggOpus, kFlac, kLinear16, kMp3 }; const char* speech_encoding_name(SpeechEncoding encoding); // Google enum name. const char* speech_encoding_mime(SpeechEncoding encoding); // Content type. // Maps a browser/content type ("audio/webm;codecs=opus", "audio/ogg", ...) // to an encoding. Returns false when unknown. bool speech_encoding_from_mime(const std::string& mime, SpeechEncoding& out); struct RecognitionRequest { std::string audio; // Raw bytes. SpeechEncoding encoding = SpeechEncoding::kWebmOpus; std::string language = "nl-NL"; int sample_rate_hz = 0; // 0: taken from the container header. bool punctuation = true; }; struct Transcript { std::string text; double confidence = 0.0; // 0 when the backend does not report one. }; class Synthesizer { public: virtual ~Synthesizer() = default; virtual Audio synthesize(const SynthesisRequest& request) const = 0; }; class Recognizer { public: virtual ~Recognizer() = default; virtual Transcript recognize(const RecognitionRequest& request) const = 0; }; } // namespace molo::google