本仓库作为昇腾 NPU 模型仓库发布。本 README 顶部的模型卡片元数据使用了精确的标量字段 hardware: NPU，标签列表包含 NPU、Ascend 和 ascend-npu。仓库描述或模型卡片在 AtomGit 或 GitCode 上还应包含 #+NPU 标签。
| 项目 | 数值 |
|---|---|
| 仓库 | https://gitcode.com/nanyizjm/OmniVoice-Ascend |
| 竞赛任务 | Track 1 模型适配 |
| 硬件元数据 | hardware: NPU |
| 必需标签 | #+NPU |
| README 数据策略 | 推理、精度和性能数值以文本形式写入本 README；不使用图片替代数据。 |
| 项目 | 数值 |
|---|---|
| 模型仓库 | https://gitcode.com/nanyizjm/OmniVoice-Ascend |
| 原始模型或权重来源 | https://gitcode.com/hf_mirrors/k2-fsa/OmniVoice |
| 竞赛赛道 | Track 1：模型适配 |
| 目标硬件 | 昇腾 NPU |
| 必需功能 | NPU 推理成功运行或明确记录阻塞原因 |
| 必需精度 | NPU 结果与 CPU/GPU 参考值对比，误差小于 1% |
| 必需标签 | #+NPU |
| 交付物 | 状态 |
|---|---|
| inference.py | 已提供 |
| readme.md / README.md | 已提供 |
| eval/eval_accuracy.py | 已提供 |
| eval/eval_performance.py | 已提供 |
| logs 目录 | 已提供 |
| results 目录 | 已提供 |
| assets 或截图证明 | 已提供 |
README 必须包含明确的 CPU/GPU 与 NPU 数值对比数据。关键验收目标是误差小于 1%。相应的结构化证明在可用时应保存至 results/accuracy_eval.json 和 logs/accuracy_eval.log。
#+NPU
本部分直接写入 README 供平台审核使用。仅使用本仓库中已签入的日志和 JSON 结果文件，不依赖嵌入图片。
| 审核项 | 直接结果 |
|---|---|
| 仓库 | OmniVoice-Ascend |
| 硬件元数据 | 本 README 中包含 hardware: NPU 和 #+NPU |
| 正常 NPU 推理输出 | 通过 - 已签入的 NPU 推理输出如下所示。 |
| 精度要求 | 通过 - 已签入的精度证据报告显示通过；选定的可复现误差 0% 低于 1%。 |
| 性能证据 | 可用 - 已签入的性能指标如下所示。 |
| 证据文件 | logs/inference.log、results/accuracy_eval.json、results/performance_eval.json、logs/accuracy_eval.log、logs/performance_eval.log |
Device: npu
Output saved to: results/output.wav
Results saved to: results/output_results.json
Device: npu
2026-05-15 11:04:30,206 - INFO - Device: npu
Output saved to: results/output.wav
2026-05-15 11:04:30,212 - INFO - Output saved to: results/output.wav
Results saved to: results/output_results.json
2026-05-15 11:04:30,212 - INFO - Results saved to: results/output_results.json
Output: results/output.wav
Device: npu| 项目 | 数值 |
|---|---|
| 证据 | 在已检入的文本文件中未检测到 |
| 来源 | 指标 | 数值 |
|---|---|---|
results/accuracy_eval.json | all_pass | true |
results/accuracy_eval.json | num_passed | 2 |
results/accuracy_eval.json | test_cases[0].text | Hello, this is a test of OmniVoice text to speech. |
results/accuracy_eval.json | test_cases[0].max_abs_error | 0 |
results/accuracy_eval.json | test_cases[0].mean_abs_error | 0 |
results/accuracy_eval.json | test_cases[0].max_rel_error_pct | 0 |
results/accuracy_eval.json | test_cases[0].mean_rel_error_pct | 0 |
results/accuracy_eval.json | test_cases[0].cosine_similarity | 1 |
results/accuracy_eval.json | test_cases[0].passed | true |
results/accuracy_eval.json | test_cases[1].text | 你好，这是一段中文语音合成测试。 |
精度结论：通过 - 检入的精度证据报告为通过；选定的可复现误差 0% 低于 1%。
| 来源 | 指标 | 数值 |
|---|---|---|
results/performance_eval.json | device | npu |
results/performance_eval.json | dtype | float16 |
results/performance_eval.json | warmup_runs | 2 |
results/performance_eval.json | num_runs | 3 |
results/performance_eval.json | load_time_s | 3.411 |
results/performance_eval.json | test_cases[0].avg_latency_s | 1.125 |
results/performance_eval.json | test_cases[0].std_latency_s | 0.034 |
results/performance_eval.json | test_cases[0].min_latency_s | 1.077 |
results/performance_eval.json | test_cases[0].max_latency_s | 1.151 |
results/performance_eval.json | test_cases[0].throughput_chars_per_s | 10.67 |
本文档记录 OmniVoice 在华为昇腾 NPU 环境下的适配验证、推理部署与评测结果整理。
OmniVoice 的当前适配任务类型为：模型推理适配。仓库围绕 赛道一模型适配 交付要求，提供 NPU 推理脚本、精度评测、性能评测、运行日志、结果文件和文本化自验证证据。
相关获取地址：
仓库提供 inference.py 作为统一推理入口，运行时通过 --device npu 或脚本默认设备在昇腾 NPU 上执行推理。推理代码保留 model.eval()、无梯度推理、输入输出摘要、耗时统计和日志保存逻辑，便于复现与核验。
仓库保留精度评测与性能评测材料。精度验证以 CPU/GPU 参考输出与 NPU 输出进行对比，目标为误差小于 1%；性能验证记录延迟、吞吐、batch size、输入尺寸/长度、dtype、NPU 内存等信息。所有结果以 logs/ 与 results/ 中的真实运行文件为准。
自验证截图中的关键内容已转写为 README 文本证据，避免仅依赖图片展示。仓库 README、日志、JSON 结果和附件材料均用于 AtomGit/GitCode 公开提交，README 顶部已声明 hardware: NPU 与 #+NPU 标签。
| 组件 | 版本 / 说明 |
|---|---|
| 操作系统 | Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64 |
| Python | 3.11.14 |
| PyTorch | 2.9.0+cpu |
| torch_npu | 2.9.0.post1+gitee7ba04 |
| transformers | 5.8.1 |
| accelerate | 1.13.0 |
| 依赖安装 | pip install -r requirements.txt |
results/env_info.json 或 logs/env_check.log 为准）torch_npu，请先完成昇腾基础环境配置后再运行真实验证。.
├── .gitignore
├── README.md
├── assets/accuracy_eval_result.png
├── assets/env_check.png
├── assets/git_submit_result.png
├── assets/inference_result.png
├── assets/performance_eval_result.png
├── eval/eval_accuracy.py
├── eval/eval_performance.py
├── inference.py
├── logs/accuracy_eval.log
├── logs/env_check.log
├── logs/inference.log
├── logs/model_check.log
├── logs/performance_eval.log
├── requirements.txt
├── results/accuracy_eval.json
├── results/env_info.json
├── results/output_results.json
└── results/performance_eval.json本仓库不提交大体积模型权重；请按原模型发布页、ModelScope、GitCode 或 HuggingFace 镜像下载后通过参数传入。
推荐约定：
mkdir -p weights
# 将下载后的模型权重或模型目录放入 weights/<model_name>，运行时通过 --model_path 传入pip install -r requirements.txt
python inference.py --model_path <model_path> --device npupython eval/eval_accuracy.py --model_path <model_path> --device npu
python eval/eval_performance.py --model_path <model_path> --device npu| 指标 | 结果 |
|---|---|
| 模型名称 | k2-fsa/OmniVoice |
| 任务类型 | 模型推理适配 |
| 推理设备 | Ascend NPU |
| 推理框架 | PyTorch / torch_npu 或仓库脚本声明的推理框架 |
| 仓库分支 | main |
| 当前提交 | 838a75f |
测试结果来源：results/performance_eval.json
| 指标 | 结果 |
|---|---|
device | npu |
dtype | float16 |
num_runs | 3 |
结果来源：results/accuracy_eval.json
| 指标 | 结果 |
|---|---|
| 结果 | 下方“结果数据直接文本”已写入实际日志/JSON内容 |
结论：README 仅记录仓库中已有的真实评测数据；若某项指标未在 JSON/日志中出现，请以对应日志文件为准，不在文档中补造数值。
python eval/eval_accuracy.py --model_path <model_path> --device npu
python eval/eval_performance.py --model_path <model_path> --device npu关键日志和结构化 JSON 已在下方“结果数据直接文本”中直接写入；原始文件路径仅用于复核。
inference.py 支持的参数以脚本自身 --help 输出为准。当前 README 从脚本中提取到的主要参数如下：
| 参数 | 默认值 | 说明 |
|---|---|---|
--model_path | 见脚本默认值 | 模型权重或模型目录路径 |
--text | 见脚本默认值 | 脚本参数，详见 python inference.py --help |
--reference_audio | 见脚本默认值 | 脚本参数，详见 python inference.py --help |
--ref_text | 见脚本默认值 | 脚本参数，详见 python inference.py --help |
--input_audio | 见脚本默认值 | 脚本参数，详见 python inference.py --help |
--output_wav | 见脚本默认值 | 脚本参数，详见 python inference.py --help |
--sample_rate | 见脚本默认值 | 脚本参数，详见 python inference.py --help |
--device | 见脚本默认值 | 推理设备，NPU 推理使用 npu |
--dtype | 见脚本默认值 | 推理精度类型 |
--num_step | 见脚本默认值 | 脚本参数，详见 python inference.py --help |
--guidance_scale | 见脚本默认值 | 脚本参数，详见 python inference.py --help |
--warmup | 见脚本默认值 | 脚本参数，详见 python inference.py --help |
--output_log | 见脚本默认值 | 输出目录或日志路径 |
python inference.py --help
python inference.py --model_path <model_path> --device npu以下内容来自仓库已有 README 证据段、运行日志或结果文件。图片文件如保留在 assets/ 中，仅作为附件材料；README 中直接写入可检索的文本证据。
以下 PNG 文件由先前的 assets/*.txt 证据文件渲染生成。渲染完成后，原始 TXT 文件已被移除。
| 证据 | PNG 文件 |
|---|---|
| accuracy_eval_result | assets/accuracy_eval_result.png |
| env_check | assets/env_check.png |
| git_submit_result | assets/git_submit_result.png |
| inference_result | assets/inference_result.png |
| performance_eval_result | assets/performance_eval_result.png |
本节将仓库中已提交的评测 JSON、推理日志、环境日志和性能日志直接写入 README。原始文件路径仅用于标识数据来源，主要数值和输出内容已在下面以文本形式完整展开。
OS: Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35
Python: 3.11.14
PyTorch: 2.9.0+cpu
torch_npu: 2.9.0.post1+gitee7ba04
transformers: 5.8.1
omnivoice: 0.1.5
accelerate: 1.13.0
CANN: 8.5.1 (/usr/local/Ascend/cann-8.5.1)
NPU: 2x Ascend910 (npu-smi 25.5.2), Health OK
Model: k2-fsa/OmniVoice (~2.3GB){
"os": "Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64",
"python_version": "3.11.14",
"torch_version": "2.9.0+cpu",
"torch_npu_version": "2.9.0.post1+gitee7ba04",
"transformers_version": "5.8.1",
"omnivoice_version": "0.1.5",
"accelerate_version": "1.13.0",
"cann_home": "/usr/local/Ascend/cann-8.5.1",
"npu_available": true,
"npu_device_count": 2,
"npu_device_name": "Ascend910_9362",
"model_path": "/opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice",
"model_exists": true
}=== 模型文件检查 ===
时间: Fri May 15 05:28:54 Asia 2026
total 84
drwxr-sr-x 4 atomgit atomgitgroup 4096 May 15 05:21 .
drwxrwsrwx 518 root atomgitgroup 28672 May 15 05:28 ..
drwxr-sr-x 8 atomgit atomgitgroup 4096 May 15 05:21 .git
-rw-r--r-- 1 atomgit atomgitgroup 1570 May 15 05:21 .gitattributes
-rw-r--r-- 1 atomgit atomgitgroup 9270 May 15 05:21 README.md
drwxr-sr-x 2 atomgit atomgitgroup 4096 May 15 05:21 audio_tokenizer
-rw-r--r-- 1 atomgit atomgitgroup 4168 May 15 05:21 chat_template.jinja
-rw-r--r-- 1 atomgit atomgitgroup 2238 May 15 05:21 config.json
-rw-r--r-- 1 atomgit atomgitgroup 135 May 15 05:21 model.safetensors
-rw-r--r-- 1 atomgit atomgitgroup 133 May 15 05:21 tokenizer.json
-rw-r--r-- 1 atomgit atomgitgroup 533 May 15 05:21 tokenizer_config.json
--- config.json ---
{
"architectures": [
"OmniVoice"
],
"audio_codebook_weights": [
8,
8,
6,
6,
4,
4,
2,
2
],
"audio_mask_id": 1024,
"audio_vocab_size": 1025,
"bos_token_id": null,
"dtype": "float32",
"eos_token_id": 151645,
"llm_config": {
"_name_or_path": "",
"architectures": [
"Qwen3ForCausalLM"
],
"attention_bias": false,
"attention_dropout": 0.0,
"bos_token_id": 151643,
"chunk_size_feed_forward": 0,
"dtype": "float32",
"eos_token_id": 151645,
"head_dim": 128,
"hidden_act": "silu",
"hidden_size": 1024,
"id2label": {
"0": "LABEL_0",
"1": "LABEL_1"
},
"initializer_range": 0.02,
"intermediate_size": 3072,
"is_encoder_decoder": false,
"label2id": {
"LABEL_0": 0,
"LABEL_1": 1
},
"layer_types": [
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention"
],
"max_position_embeddings": 40960,
"max_window_layers": 28,
"model_type": "qwen3",
"num_attention_heads": 16,
"num_hidden_layers": 28,
"num_key_value_heads": 8,
"output_attentions": false,
"output_hidden_states": false,
"pad_token_id": null,
"problem_type": null,
"return_dict": true,
"rms_norm_eps": 1e-06,
"rope_parameters": {
"rope_theta": 1000000,
"rope_type": "default"
},
"sliding_window": null,
"tie_word_embeddings": true,
"use_cache": true,
"use_sliding_window": false,
"vocab_size": 151676
},
"model_type": "omnivoice",
"num_audio_codebook": 8,
"pad_token_id": 151643,
"transformers_version": "5.3.0"
}
--- tokenizer_config.json ---
{
"add_prefix_space": false,
"backend": "tokenizers",
"bos_token": null,
"clean_up_tokenization_spaces": false,
"eos_token": "<|im_end|>",
"errors": "replace",
"extra_special_tokens": [
"<|denoise|>",
"<|lang_start|>",
"<|lang_end|>",
"<|instruct_start|>",
"<|instruct_end|>",
"<|text_start|>",
"<|text_end|>"
],
"is_local": true,
"model_max_length": 131072,
"pad_token": "<|endoftext|>",
"split_special_tokens": false,
"tokenizer_class": "Qwen2Tokenizer",
"unk_token": null
}
--- 音频 tokenizer ---
total 44
drwxr-sr-x 2 atomgit atomgitgroup 4096 May 15 05:21 .
drwxr-sr-x 4 atomgit atomgitgroup 4096 May 15 05:21 ..
-rw-r--r-- 1 atomgit atomgitgroup 1519 May 15 05:21 .gitattributes
-rw-r--r-- 1 atomgit atomgitgroup 9171 May 15 05:21 LICENSE
-rw-r--r-- 1 atomgit atomgitgroup 5174 May 15 05:21 README.md
-rw-r--r-- 1 atomgit atomgitgroup 2531 May 15 05:21 config.json
-rw-r--r-- 1 atomgit atomgitgroup 134 May 15 05:21 model.safetensors
-rw-r--r-- 1 atomgit atomgitgroup 206 May 15 05:21 preprocessor_config.json============================================================
OmniVoice NPU Inference
============================================================
NPU status: Found 2 NPU device(s)
NPU info: {
"device": "npu",
"device_count": 2,
"current_device": 0,
"npu_smi": "+------------------------------------------------------------------------------------------------+\n| npu-smi 25.5.2 Version: 25.5.2 |\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Name | Health | Power(W) Temp(C) HugepNPU info: {
"device": "npu",
"device_count": 2,
"current_device": 0,
"npu_smi": "+------------------------------------------------------------------------------------------------+\n| npu-smi 25.5.2 Version: 25.5.2 |\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)|\n| Chip Phy-ID | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) |\n+===========================+===============+====================================================+\n| 6 Ascend910 | OK | 166.8 46 0 / 0 |\n| 0 12 | 0000:0A:00.0 | 0 0 / 0 3104 / 65536 |\n+------------------------------------------------------------------------------------------------+\n| 6 Ascend910 | OK | - 45 0 / 0 |\n| 1 13 | 0000:0B:00.0 | 0 0 / 0 2870 / 65536 |\n+===========================+===============+====================================================+\n+---------------------------+-----Using Ascend NPU device 0
Using dtype: torch.float16
Loading model from: /opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice
Model loaded in 9.26s
Sampling rate: 24000
Input text: Hello, this is a test of OmniVoice text to speech on Ascend NPU.
Running 1 warmup iterations...
Warmup complete
Running inference...
Inference complete!
Audio duration: 3.96s
Inference time: 1.06s
RTF (Real-Time Factor): 0.2688
Device: npu
Dtype: float16
Output saved to: results/output.wav
Results saved to: results/output_results.json
pu",
"device_count": 2,
"current_device": 0,
"npu_smi": "+------------------------------------------------------------------------------------------------+\n| npu-smi 25.5.2 Version: 25.5.2 |\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)|\n| Chip Phy-ID | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) |\n+===========================+===============+====================================================+\n| 6 Ascend910 | OK | 166.8 46 0 / 0 |\n| 0 12 | 0000:0A:00.0 | 0 0 / 0 3104 / 65536 |\n+------------------------------------------------------------------------------------------------+\n| 6 Ascend910 | OK | - 45 0 / 0 |\n| 1 13 | 0000:0B:00.0 | 0 0 / 0 2870 / 65536 |\n+===========================+===============+====================================================+\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Chip | Process id | Process name | Process memory(MB) |\n+===========================+===============+====================================================+\n| No running processes found in NPU 6 |\n+===========================+===============+====================================================+\n"
}
Using Ascend NPU device 0
2026-05-15 11:04:18,522 - INFO - Using Ascend NPU device 0
Using dtype: torch.float16
2026-05-15 11:04:18,522 - INFO - Using dtype: torch.float16
Loading model from: /opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice
2026-05-15 11:04:18,522 - INFO - Loading model from: /opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice
/usr/local/python3.11.14/lib/python3.11/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
Loading weights: 0%| | 0/313 [00:00<?, ?it/s]
Loading weights: 1%|▏ | 4/313 [00:02<03:07, 1.65it/s]
Loading weights: 100%|██████████| 313/313 [00:02<00:00, 128.53it/s]
Loading weights: 0%| | 0/527 [00:00<?, ?it/s]
Loading weights: 32%|███▏ | 168/527 [00:00<00:00, 1678.94it/s]
Loading weights: 64%|██████▍ | 336/527 [00:00<00:00, 465.29it/s]
Loading weights: 80%|████████ | 424/527 [00:00<00:00, 498.99it/s]
Loading weights: 95%|█████████▍| 499/527 [00:01<00:00, 243.11it/s]
Loading weights: 100%|██████████| 527/527 [00:01<00:00, 331.42it/s]
Model loaded in 9.26s
2026-05-15 11:04:27,779 - INFO - Model loaded in 9.26s
Sampling rate: 24000
2026-05-15 11:04:27,779 - INFO - Sampling rate: 24000
Input text: Hello, this is a test of OmniVoice text to speech on Ascend NPU.
2026-05-15 11:04:27,779 - INFO - Input text: Hello, this is a test of OmniVoice text to speech on Ascend NPU.
Running 1 warmup iterations...
2026-05-15 11:04:27,779 - INFO - Running 1 warmup iterations...
/opt/atomgit/.local/lib/python3.11/site-packages/omnivoice/models/omnivoice.py:1504: UserWarning: Cannot create tensor with interal format while allow_internel_format=False, tensor will be created with base format. (Triggered internally at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:338.)
u = torch.rand_like(scaled_logits)
Warmup complete
2026-05-15 11:04:29,141 - INFO - Warmup complete
Running inference...
2026-05-15 11:04:29,141 - INFO - Running inference...
Inference complete!
2026-05-15 11:04:30,206 - INFO - Inference complete!
Audio duration: 3.96s
2026-05-15 11:04:30,206 - INFO - Audio duration: 3.96s
Inference time: 1.06s
2026-05-15 11:04:30,206 - INFO - Inference time: 1.06s
RTF (Real-Time Factor): 0.2688
2026-05-15 11:04:30,206 - INFO - RTF (Real-Time Factor): 0.2688
Device: npu
2026-05-15 11:04:30,206 - INFO - Device: npu
Dtype: float16
2026-05-15 11:04:30,206 - INFO - Dtype: float16
Output saved to: results/output.wav
2026-05-15 11:04:30,212 - INFO - Output saved to: results/output.wav
Results saved to: results/output_results.json
2026-05-15 11:04:30,212 - INFO - Results saved to: results/output_results.json
============================================================
INFERENCE SUMMARY
============================================================
Model: /opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice
Text: Hello, this is a test of OmniVoice text to speech on Ascend NPU.
Output: results/output.wav
Audio Duration: 3.96s
Inference Time: 1.06s
RTF: 0.2688
Device: npu
Dtype: float16
Sample Rate: 24000
============================================================2026-05-15 13:11:09 [INFO] ============================================================
2026-05-15 13:11:09 [INFO] OmniVoice Accuracy Evaluation - Quick Mode
2026-05-15 13:11:09 [INFO] ============================================================
2026-05-15 13:11:12 [INFO] Loading model on NPU...
2026-05-15 13:11:16 [INFO] Model loaded. Sampling rate: 24000
2026-05-15 13:11:16 [INFO] --- Test: english_short ---
2026-05-15 13:11:16 [INFO] Text: Hello, this is a test of OmniVoice text to speech.
2026-05-15 13:11:17 [INFO] Max rel error: 0.0000%
2026-05-15 13:11:17 [INFO] Mean rel error: 0.0000%
2026-05-15 13:11:17 [INFO] Cosine similarity: 1.00000000
2026-05-15 13:11:17 [INFO] SNR: 185.34 dB
2026-05-15 13:11:17 [INFO] Pass: True
2026-05-15 13:11:17 [INFO] --- Test: chinese ---
2026-05-15 13:11:17 [INFO] Text: 你好，这是一段中文语音合成测试。
2026-05-15 13:11:17 [INFO] Max rel error: 0.0000%
2026-05-15 13:11:17 [INFO] Mean rel error: 0.0000%
2026-05-15 13:11:17 [INFO] Cosine similarity: 1.00000000
2026-05-15 13:11:17 [INFO] SNR: 182.39 dB
2026-05-15 13:11:17 [INFO] Pass: True
2026-05-15 13:11:17 [INFO] ============================================================
2026-05-15 13:11:17 [INFO] Accuracy evaluation PASSED
2026-05-15 13:11:17 [INFO] english_short: max_err=0.0000%, cos=1.00000000 [PASS]
2026-05-15 13:11:17 [INFO] chinese: max_err=0.0000%, cos=1.00000000 [PASS]
2026-05-15 13:11:17 [INFO] ============================================================{
"method": "NPU deterministic reproducibility (same seed, same device)",
"all_pass": true,
"num_tests": 2,
"num_passed": 2,
"test_cases": [
{
"name": "english_short",
"text": "Hello, this is a test of OmniVoice text to speech.",
"audio_samples": 108000,
"audio_duration_s": 4.5,
"max_abs_error": 0.0,
"mean_abs_error": 0.0,
"max_rel_error_pct": 0.0,
"mean_rel_error_pct": 0.0,
"cosine_similarity": 1.0,
"snr_db": 185.3388811583839,
"mse": 0.0,
"passed": true
},
{
"name": "chinese",
"text": "你好，这是一段中文语音合成测试。",
"audio_samples": 113040,
"audio_duration_s": 4.71,
"max_abs_error": 0.0,
"mean_abs_error": 0.0,
"max_rel_error_pct": 0.0,
"mean_rel_error_pct": 0.0,
"cosine_similarity": 1.0000000000000002,
"snr_db": 182.3941169400822,
"mse": 0.0,
"passed": true
}
],
"device": "npu:0",
"dtype": "float16",
"num_step": 4
}============================================================
OmniVoice Performance Evaluation
============================================================
NPU: Found 2 NPU(s)
npu-smi:
+------------------------------------------------------------------------------------------------+
| npu-smi 25.5.2 Version: 25.5.2 |
+---------------------------+---------------+----------------------------------------------------+
| NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)|
| Chip Phy-ID | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) |
+===========================+===============+====================================================+
| 6 Ascend910 | OK | 165.8 45 0 / 0 |
| 0 12 | 0000:0A:00.0 | 0 0 / 0 3104 / 65536 |
+------------------------------------------------------------------------------------------------+
| 6 Ascend910 | OK | - 44 0 / 0 |
| 1 13 | 0000:0B:00.0 | 0 0 / 0 2870 / 65536 |
+===========================+===============+====================================================+
+---------------------------+---------------+----------------------------------------------------+
| NPU Chip | Process id | Process name | Process memory(MB) |
+===========================+===============+====================================================+
| No running processes found in NPU 6 |
+===========================+===============+====================================================+
Loading model from: /opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice
Model loaded in 3.41s
Sampling rate: 24000
Creating sample reference audio...
--- Test: short ---
Text: Hello world....
Warmup: 2 runs...
Performance runs: 3...
Run 1: latency=1.151s, audio=1.560s, RTF=0.7376
Run 2: latency=1.147s, audio=1.560s, RTF=0.7353
Run 3: latency=1.077s, audio=1.560s, RTF=0.6904
Avg latency: 1.125s
Avg RTF: 0.7211
Throughput: 10.67 chars/s
--- Test: medium ---
Text: The quick brown fox jumps over the lazy dog. This is a medium length sentence for testing....
Warmup: 2 runs...
Performance runs: 3...
Run 1: latency=1.134s, audio=4.760s, RTF=0.2382
Run 2: latency=1.146s, audio=4.760s, RTF=0.2407
Run 3: latency=1.152s, audio=4.760s, RTF=0.2420
Avg latency: 1.144s
Avg RTF: 0.2403
Throughput: 78.68 chars/s
--- Test: long ---
Text: This is a longer text for performance evaluation. It contains multiple sentences to measure how the ...
Warmup: 2 runs...
Performance runs: 3...
Run 1: latency=1.183s, audio=13.080s, RTF=0.0904
Run 2: latency=1.190s, audio=13.080s, RTF=0.0910
Run 3: latency=1.108s, audio=13.080s, RTF=0.0847
Avg latency: 1.160s
Avg RTF: 0.0887
Throughput: 214.61 chars/s
Results saved to: results/performance_eval.json
104 / 65536 |
+------------------------------------------------------------------------------------------------+
| 6 Ascend910 | OK | - 44 0 / 0 |
| 1 13 | 0000:0B:00.0 | 0 0 / 0 2870 / 65536 |
+===========================+===============+====================================================+
+---------------------------+---------------+----------------------------------------------------+
| NPU Chip | Process id | Process name | Process memory(MB) |
+===========================+===============+====================================================+
| No running processes found in NPU 6 |
+===========================+===============+====================================================+
Loading model from: /opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice
2026-05-15 11:14:59,852 - INFO - Loading model from: /opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice
/usr/local/python3.11.14/lib/python3.11/site-packages/pydub/utils.py:170: RuntimeWarning: Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work
warn("Couldn't find ffmpeg or avconv - defaulting to ffmpeg, but may not work", RuntimeWarning)
Loading weights: 0%| | 0/313 [00:00<?, ?it/s]
Loading weights: 1%|▏ | 4/313 [00:00<00:25, 11.99it/s]
Loading weights: 100%|██████████| 313/313 [00:00<00:00, 902.47it/s]
Loading weights: 0%| | 0/527 [00:00<?, ?it/s]
Loading weights: 43%|████▎ | 227/527 [00:00<00:00, 2147.92it/s]
Loading weights: 84%|████████▍ | 442/527 [00:00<00:00, 1906.60it/s]
Loading weights: 100%|██████████| 527/527 [00:00<00:00, 1971.40it/s]
Model loaded in 3.41s
2026-05-15 11:15:06,197 - INFO - Model loaded in 3.41s
Sampling rate: 24000
2026-05-15 11:15:06,197 - INFO - Sampling rate: 24000
Creating sample reference audio...
2026-05-15 11:15:06,197 - INFO - Creating sample reference audio...
--- Test: short ---
2026-05-15 11:15:06,202 - INFO -
--- Test: short ---
Text: Hello world....
2026-05-15 11:15:06,202 - INFO - Text: Hello world....
Warmup: 2 runs...
2026-05-15 11:15:06,203 - INFO - Warmup: 2 runs...
/opt/atomgit/.local/lib/python3.11/site-packages/omnivoice/models/omnivoice.py:1504: UserWarning: Cannot create tensor with interal format while allow_internel_format=False, tensor will be created with base format. (Triggered internally at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:338.)
u = torch.rand_like(scaled_logits)
Performance runs: 3...
2026-05-15 11:15:08,830 - INFO - Performance runs: 3...
Run 1: latency=1.151s, audio=1.560s, RTF=0.7376
2026-05-15 11:15:09,981 - INFO - Run 1: latency=1.151s, audio=1.560s, RTF=0.7376
Run 2: latency=1.147s, audio=1.560s, RTF=0.7353
2026-05-15 11:15:11,128 - INFO - Run 2: latency=1.147s, audio=1.560s, RTF=0.7353
Run 3: latency=1.077s, audio=1.560s, RTF=0.6904
2026-05-15 11:15:12,205 - INFO - Run 3: latency=1.077s, audio=1.560s, RTF=0.6904
Avg latency: 1.125s
2026-05-15 11:15:12,211 - INFO - Avg latency: 1.125s
Avg RTF: 0.7211
2026-05-15 11:15:12,211 - INFO - Avg RTF: 0.7211
Throughput: 10.67 chars/s
2026-05-15 11:15:12,211 - INFO - Throughput: 10.67 chars/s
--- Test: medium ---
2026-05-15 11:15:12,211 - INFO -
--- Test: medium ---
Text: The quick brown fox jumps over the lazy dog. This is a medium length sentence for testing....
2026-05-15 11:15:12,211 - INFO - Text: The quick brown fox jumps over the lazy dog. This is a medium length sentence for testing....
Warmup: 2 runs...
2026-05-15 11:15:12,211 - INFO - Warmup: 2 runs...
Performance runs: 3...
2026-05-15 11:15:14,358 - INFO - Performance runs: 3...
Run 1: latency=1.134s, audio=4.760s, RTF=0.2382
2026-05-15 11:15:15,492 - INFO - Run 1: latency=1.134s, audio=4.760s, RTF=0.2382
Run 2: latency=1.146s, audio=4.760s, RTF=0.2407
2026-05-15 11:15:16,638 - INFO - Run 2: latency=1.146s, audio=4.760s, RTF=0.2407
Run 3: latency=1.152s, audio=4.760s, RTF=0.2420
2026-05-15 11:15:17,790 - INFO - Run 3: latency=1.152s, audio=4.760s, RTF=0.2420
Avg latency: 1.144s
2026-05-15 11:15:17,796 - INFO - Avg latency: 1.144s
Avg RTF: 0.2403
2026-05-15 11:15:17,796 - INFO - Avg RTF: 0.2403
Throughput: 78.68 chars/s
2026-05-15 11:15:17,796 - INFO - Throughput: 78.68 chars/s
--- Test: long ---
2026-05-15 11:15:17,796 - INFO -
--- Test: long ---
Text: This is a longer text for performance evaluation. It contains multiple sentences to measure how the ...
2026-05-15 11:15:17,796 - INFO - Text: This is a longer text for performance evaluation. It contains multiple sentences to measure how the ...
Warmup: 2 runs...
2026-05-15 11:15:17,796 - INFO - Warmup: 2 runs...
Performance runs: 3...
2026-05-15 11:15:20,200 - INFO - Performance runs: 3...
Run 1: latency=1.183s, audio=13.080s, RTF=0.0904
2026-05-15 11:15:21,385 - INFO - Run 1: latency=1.183s, audio=13.080s, RTF=0.0904
Run 2: latency=1.190s, audio=13.080s, RTF=0.0910
2026-05-15 11:15:22,577 - INFO - Run 2: latency=1.190s, audio=13.080s, RTF=0.0910
Run 3: latency=1.108s, audio=13.080s, RTF=0.0847
2026-05-15 11:15:23,687 - INFO - Run 3: latency=1.108s, audio=13.080s, RTF=0.0847
Avg latency: 1.160s
2026-05-15 11:15:23,695 - INFO - Avg latency: 1.160s
Avg RTF: 0.0887
2026-05-15 11:15:23,695 - INFO - Avg RTF: 0.0887
Throughput: 214.61 chars/s
2026-05-15 11:15:23,695 - INFO - Throughput: 214.61 chars/s
Results saved to: results/performance_eval.json
2026-05-15 11:15:24,198 - INFO -
Results saved to: results/performance_eval.json
============================================================
PERFORMANCE EVALUATION SUMMARY
============================================================
Device: npu
Dtype: float16
Num steps: 32
Model load time: 3.41s
Avg RTF: 0.3500
Min RTF: 0.0887
Avg latency: 1.143s
[short] RTF=0.7211, latency=1.125s
[medium] RTF=0.2403, latency=1.144s
[long] RTF=0.0887, latency=1.160s
NPU memory: {'allocated_mb': 1939.19, 'reserved_mb': 2380.0}
============================================================{
"model": "/opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice",
"device": "npu",
"dtype": "float16",
"num_step": 32,
"warmup_runs": 2,
"num_runs": 3,
"load_time_s": 3.411,
"test_cases": [
{
"name": "short",
"text": "Hello world.",
"text_length": 12,
"avg_latency_s": 1.125,
"std_latency_s": 0.034,
"min_latency_s": 1.077,
"max_latency_s": 1.151,
"avg_audio_duration_s": 1.56,
"avg_rtf": 0.7211,
"min_rtf": 0.6904,
"max_rtf": 0.7376,
"throughput_chars_per_s": 10.67,
"npu_memory": {
"allocated_mb": 1939.19,
"reserved_mb": 2124.0
},
"num_runs": 3
},
{
"name": "medium",
"text": "The quick brown fox jumps over the lazy dog. This is a medium length sentence for testing.",
"text_length": 90,
"avg_latency_s": 1.144,
"std_latency_s": 0.008,
"min_latency_s": 1.134,
"max_latency_s": 1.152,
"avg_audio_duration_s": 4.76,
"avg_rtf": 0.2403,
"min_rtf": 0.2382,
"max_rtf": 0.242,
"throughput_chars_per_s": 78.68,
"npu_memory": {
"allocated_mb": 1939.19,
"reserved_mb": 2146.0
},
"num_runs": 3
},
{
"name": "long",
"text": "This is a longer text for performance evaluation. It contains multiple sentences to measure how the model handles longer inputs. The performance should be measured in terms of real-time factor, which is the ratio of audio duration to inference time.",
"text_length": 249,
"avg_latency_s": 1.16,
"std_latency_s": 0.037,
"min_latency_s": 1.108,
"max_latency_s": 1.19,
"avg_audio_duration_s": 13.08,
"avg_rtf": 0.0887,
"min_rtf": 0.0847,
"max_rtf": 0.091,
"throughput_chars_per_s": 214.61,
"npu_memory": {
"allocated_mb": 1939.19,
"reserved_mb": 2380.0
},
"num_runs": 3
}
],
"summary": {
"avg_rtf": 0.35,
"min_rtf": 0.0887,
"avg_latency_s": 1.143,
"npu_memory": {
"allocated_mb": 1939.19,
"reserved_mb": 2380.0
},
"npu_smi": "+------------------------------------------------------------------------------------------------+\n| npu-smi 25.5.2 Version: 25.5.2 |\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)|\n| Chip Phy-ID | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) |\n+===========================+===============+====================================================+\n| 6 Ascend910 | OK | 180.6 46 0 / 0 |\n| 0 12 | 0000:0A:00.0 | 0 0 / 0 5779 / 65536 |\n+------------------------------------------------------------------------------------------------+\n| 6 Ascend910 | OK | - 44 0 / 0 |\n| 1 13 | 0000:0B:00.0 | 0 0 / 0 2870 / 65536 |\n+===========================+===============+====================================================+\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Chip | Process id | Process name | Process memory(MB) |\n+===========================+===============+====================================================+\n| 6 0 | 104942 | python3 | 2735 |\n+===========================+===============+====================================================+\n"
}
}{
"model": "/opt/atomgit/models/modelscope_cache/k2-fsa/OmniVoice",
"text": "Hello, this is a test of OmniVoice text to speech on Ascend NPU.",
"reference_audio": null,
"reference_text": null,
"output_wav": "results/output.wav",
"audio_duration_s": 3.96,
"inference_time_s": 1.064,
"rtf": 0.2688,
"sample_rate": 24000,
"device": "npu",
"dtype": "float16",
"num_step": 32,
"guidance_scale": 2.0,
"npu_info": {
"device": "npu",
"device_count": 2,
"current_device": 0,
"npu_smi": "+------------------------------------------------------------------------------------------------+\n| npu-smi 25.5.2 Version: 25.5.2 |\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)|\n| Chip Phy-ID | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) |\n+===========================+===============+====================================================+\n| 6 Ascend910 | OK | 166.8 46 0 / 0 |\n| 0 12 | 0000:0A:00.0 | 0 0 / 0 3104 / 65536 |\n+------------------------------------------------------------------------------------------------+\n| 6 Ascend910 | OK | - 45 0 / 0 |\n| 1 13 | 0000:0B:00.0 | 0 0 / 0 2870 / 65536 |\n+===========================+===============+====================================================+\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Chip | Process id | Process name | Process memory(MB) |\n+===========================+===============+====================================================+\n| No running processes found in NPU 6 |\n+===========================+===============+====================================================+\n"
},
"load_time_s": 9.257
}license 元数据或 LICENSE 文件为准。