c++ runtime 2pass 循环解码同一句话，会使得offline模型解码越来越慢吗？

[locasxe]

❓ Questions and Help

What is your question?

运行 runtime 源码，加入循环，解码（音频内容为 123456789），重复大概几千遍，解码的速率会越来越慢，希望可以看看什么原因

Code

demo改动如下： for (int ii = 0; ii < 10000; ii++) { # 推理的时候加入了这个循环。 for (int sample_offset = 0; sample_offset < buff_len; sample_offset += std::min(step, buff_len - sample_offset)) { if (sample_offset + step >= buff_len - 1) { step = buff_len - sample_offset; is_final = false; # 测试的时候不输入尾点 } else { is_final = false; } gettimeofday(&start, nullptr); FUNASR_RESULT result = FunTpassInferBuffer(tpass_handle, tpass_online_handle, speech_buff+sample_offset, step, punc_cache, is_final, sampling_rate_, "pcm", (ASR_TYPE)asr_mode_, hotwords_embedding, true, decoder_handle); gettimeofday(&end, nullptr); seconds = (end.tv_sec - start.tv_sec); taking_micros += ((seconds * 1000000) + end.tv_usec) - (start.tv_usec);

What's your environment?

• OS Linux:
• FunASR Version 最新:
• ModelScope Version 最新:
• PyTorch Version 2.0.0:
• How you installed funasr source:
• Python version 3.9:
• GPU T4
• CUDA/cuDNN version 12.0: 希望大佬们解答下