printf("[listening]\n");
    while (1) {
        std::unique_lock<std::mutex> lock(state.lock);
        while (state.clips.size() == 0) {
            state.cond.wait(lock);
        }
        auto clip = state.clips.front();
        state.clips.pop_front();
        lock.unlock();
        auto t_start = std::chrono::high_resolution_clock::now();
        std::vector<float> data;
        for (auto const &frame : clip) {
            std::copy(frame.samples.begin(), frame.samples.end(), std::back_inserter(data));
        }
        w2l_emission *emission = w2l_engine_process(engine, &data[0], data.size());
        char *text = NULL;
        // emit
        text = w2l_emission_text(emission);
        auto t_emit = std::chrono::high_resolution_clock::now() - t_start;
        bool emit = false;
        if (strlen(text) > 0) {
            printf("[emit] %s\n", text);
            emit = true;
        }
        free(text);
        // decode
        t_start = std::chrono::high_resolution_clock::now();
        if (decoder) {
            text = w2l_decoder_decode(decoder, emission);
            if (strlen(text) > 0)
                printf("[dec]  %s\n", text);
            free(text);
        }
        auto t_decode = std::chrono::high_resolution_clock::now() - t_start;
        w2l_emission_free(emission);

        if (emit) {
            double elapsed_emit = std::chrono::duration<double, std::milli>(t_emit).count();
            double elapsed_decode = std::chrono::duration<double, std::milli>(t_decode).count();
            double elapsed_audio = (double)clip.size() * frame_time * 1000;
            printf("[audio]=%.3fms  [emit]=%.3fms (%.2fx)  [decode]=%.3fms (%.2fx)  [total]=%.3fms (%.2fx)\n",
                elapsed_audio,
                elapsed_emit, elapsed_emit/elapsed_audio,
                elapsed_decode, elapsed_decode/elapsed_audio,
                elapsed_emit + elapsed_decode, (elapsed_emit+elapsed_decode)/elapsed_audio
            );
        }
        lock.lock();
    }