Method handle_voices

app/server/runtime.cpp:705–738

Cached-voice discovery for the "voice"/cached_voice_id request field. Families that support voice presets (e.g. pocket_tts, see assets.cpp: model_root/embeddings/ .safetensors) keep them under an "embeddings" directory next to the model weights; other families simply have no such directory and report no voices. Used by clients (llama-swap's playground, and potentially Open WebUI) that call GET

    return model.default_voice_preset.has_value() ? &*model.default_voice_preset : nullptr;
}

engine::runtime::TaskRequest ServerState::build_speech_request(const LoadedModel & model, const Value & body) const {
    engine::runtime::TaskRequest request;
    request.text_input = engine::runtime::Transcript{
        engine::io::json::require_string(body, "input"),
        engine::io::json::optional_string(body, "language", ""),
    };

    request.options = options_from_object(body.find("options"));
    add_option_from_json(request.options, body, "seed", "seed");
    add_option_from_json(request.options, body, "temperature", "temperature");
    add_option_from_json(request.options, body, "top_k", "top_k");
    add_option_from_json(request.options, body, "top_p", "top_p");
    add_option_from_json(request.options, body, "max_tokens", "max_tokens");
    add_option_from_json(request.options, body, "max_steps", "max_steps");
    add_option_from_json(request.options, body, "repetition_penalty", "repetition_penalty");
    add_option_from_json(request.options, body, "guidance_scale", "guidance_scale");
    add_option_from_json(request.options, body, "num_inference_steps", "num_inference_steps");
    if (const auto * value = body.find("instructions")) {
        request.options["instruct"] = value->as_string();
    }

    bool voice_field_is_preset = false;
    const auto * preset = select_voice_preset(model, body, voice_field_is_preset);

    engine::runtime::VoiceCondition voice;
    bool has_voice = false;
    if (preset != nullptr) {
        if (preset->voice_id.has_value()) {
            voice.speaker = engine::runtime::VoiceReference{};
            voice.speaker->cached_voice_id = *preset->voice_id;
            has_voice = true;
        }
        if (preset->audio.has_value()) {
            if (!voice.speaker.has_value()) {
                voice.speaker = engine::runtime::VoiceReference{};
            }