From 5bd6c03bf703871f11607fb97a6563045747a894 Mon Sep 17 00:00:00 2001 From: coding-lz Date: Fri, 15 Dec 2023 17:38:15 +0800 Subject: [PATCH] =?UTF-8?q?=E4=BF=AE=E6=94=B9FastSpeech=E8=84=9A=E6=9C=AC?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- AscendIE/TorchAIE/built-in/audio/FastSpeech2/README.md | 4 ++-- AscendIE/TorchAIE/built-in/audio/FastSpeech2/aie_val.py | 1 - 2 files changed, 2 insertions(+), 3 deletions(-) diff --git a/AscendIE/TorchAIE/built-in/audio/FastSpeech2/README.md b/AscendIE/TorchAIE/built-in/audio/FastSpeech2/README.md index 7a26749fa5..3fc22eb0d0 100644 --- a/AscendIE/TorchAIE/built-in/audio/FastSpeech2/README.md +++ b/AscendIE/TorchAIE/built-in/audio/FastSpeech2/README.md @@ -125,9 +125,9 @@ FastSpeech2是一种非自回归的语音合成网络。所谓自回归是指模 # 模型推理性能&精度 -调用ACL接口推理计算，性能&精度参考下列数据。 +性能&精度参考下列数据。 | 芯片型号 | Batch Size | 数据集 | 精度 | 性能 | |:-----------:|:-------------:|:-----------:|:--------:|:------------:| -| Ascend310P3 | 1 | LJSpeech | 人工判断语音质量 | 13.28 wavs/s | +| Ascend310P3 | 1 | LJSpeech | 人工判断语音质量 | 12.40 wavs/s | - 说明：由于模型推理为多个子模型串联，故性能采用端到端推理LJSpeech验证集中512条文本数据测得。 diff --git a/AscendIE/TorchAIE/built-in/audio/FastSpeech2/aie_val.py b/AscendIE/TorchAIE/built-in/audio/FastSpeech2/aie_val.py index c5d37691cf..deba963780 100644 --- a/AscendIE/TorchAIE/built-in/audio/FastSpeech2/aie_val.py +++ b/AscendIE/TorchAIE/built-in/audio/FastSpeech2/aie_val.py @@ -164,7 +164,6 @@ if __name__ == "__main__": batch_data.sort(key=lambda x : x[5]) for batch in tqdm(batch_data): cnt += 1 - batch = to_device(batch, 'cpu') src_masks = get_mask_from_lengths(batch[4], batch[5]) mel_output, mel_lens = fastspeech2_infer(batch[3], src_masks, fastspeech2_aie, control_values) hifigan_infer(batch[0], mel_output, mel_lens, preprocess_config, train_config, hifigan_aie) -- Gitee