Total runs:
1.1K
Run Growth:
-506
Growth Rate:
-41.82%
Sponsored by
APIMart
AXERA-TECH / 3D-Speaker-Meeting-Summary
Model's Last Updated: February 11 2026
audio-text-to-text
Introduction of 3D-Speaker-Meeting-Summary
Model Details of 3D-Speaker-Meeting-Summary
3D-Speaker-MT.axera
meeting transcription demo on Axera
- Python 示例
- C++ 示例
Convert tools links:
For those who are interested in model conversion, you can try to export axmodel through the original repo :
How to Convert from ONNX to axmodel
支持平台
- AX650N
功能
- 会议音频转录与总结
模型转换
参考 模型转换
上板部署
- AX650N 的设备已预装 Ubuntu22.04
- 以 root 权限登陆 AX650N 的板卡设备
- 链接互联网，确保 AX650N 的设备能正常执行 apt install, pip install 等指令
- 已验证设备：AX650N DEMO Board
Python API 运行
在python3.10(验证)
Requirements
pip3 install -r requirements.txt
在开发板运行以下命令
支持输入音频文件格式：wav，mp3
axera@dell:~/3D-Speaker-MT.axera$ python gradio_demo.py
[INFO] Available providers: ['AXCLRTExecutionProvider']
[INFO] Using provider: AXCLRTExecutionProvider
[INFO] SOC Name: AX650N
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Compiler version: 4.2 057581a9
[INFO] Using provider: AXCLRTExecutionProvider
[INFO] SOC Name: AX650N
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Compiler version: 4.2 057581a9
[INFO] Using provider: AXCLRTExecutionProvider
[INFO] SOC Name: AX650N
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Compiler version: 4.2 2cc4ac86
============================================================
🎤 AI会议转录与总结系统 - 启动成功！
============================================================
🌐 服务已启动，可通过以下链接访问：
📍 本地访问: http://127.0.0.1:7860
🏠 局域网访问: http://10.126.33.124:7860
🌍 外网访问: 运行时添加参数 share=True
============================================================
💡 提示：点击上方链接即可在浏览器中打开应用
============================================================
* Running on local URL: http://0.0.0.0:7860
* To create a public link, set `share=True` in `launch()`.
python3 ax_meeting_transc_demo.py --output_dir output_dir --wav_file wav/vad_example.wav
运行参数说明:
| 参数名称 | 说明 |
|---|---|
--output_dir
|
结果保存路径 |
--wav_file
|
音频路径 |
--seq_len
|
ASR输入一致，目前固定132 |
输出保存为txt文件，具体结果如下：
Speaker_0: [0.000 63.810] 试错的过程很简单，而且特别是今天报名仓雪卡的同学，你们可以。听到后面的有专门的活动课，他会大大降低你的试绸成本。其实你也可以不来听课。为什么你自己写嘛？我写今天写5个点，我就试试试验一下，反正这5个点不行，我再写5个点，这是再不行。那再写5个点吧，。你总会所谓的活动大神和所谓的高手都是只有一个。把所有的错，所有的坑全国趟一遍，留下正确的你就是所谓的大神。明白吗？所以说关于活动通过这一块，我只送给你们四个字啊，换位思考。如果说你要想降低。你的试错成本，今天来这里你们就是对的。。因为有畅血唱血卡这个机会，所以说关于活动过于不过这个问题，或者活动很难通过这个话题。呃，如果真的。那要坐下来聊的话，要聊一天。但是我觉得我刚才说的四个字足够。好，谢谢。
Speaker_1: [63.810 70.471] 好，非常感谢那个三茂老师的回答啊。三茂老师说我们在。整个店铺的这个活动当中，我们要学会换位思考。其实我。
Latency
AX650N
| model | latency(ms) |
|---|---|
| vad |
5.441
|
| cammplus |
2.907
|
| sensevoice |
25.482
|
RTF: 约为0.2
eg:
Inference time for vad_example.wav: 10.92 seconds
- VAD processing time: 2.20 seconds
- Speaker embedding extraction time: 1.88 seconds
- Speaker clustering time: 0.16 seconds
- ASR processing time: 3.75 seconds
load model + Inference time for vad_example.wav: 13.08 seconds
Audio duration: 70.47 seconds
RTF: 0.15
执行
demo.py
获取完整的会议记录总结:
注意: LLM 部分使用 Qwen3-4B-Int4 (36 Layers), 8K 上下文, 采用 4 + 4 模式编译 (即 4k prefill + 4k decode 的模式).
python3 demo.py --output_dir output_dir --wav_file wav/vad_example.wav
输出:
Model loaded successfully!
slice_indices: [0, 1]
Slice prefill done: 0
Slice prefill done: 1
answer >> **参会人员总结：**
会议中明确出现的发言人包括：
- **三茂老师**（也称“Speaker_0”）：主讲人，围绕“活动试错成本”“换位思考”等核心议题进行分享，提出“换位思考”是降低试错成本的关键方法，并强调“活动大神”是通过遍历所有错误后才形成的。
- **另一位与会者**（Speaker_1）：在三茂老师发言后进行补充，表示认同其观点，进一步展开“换位思考”在店铺活动中的实践意义。
---
**内容摘要：**
本次会议聚焦于**活动策划与执行中的试错成本问题**，核心观点如下：
1. **降低试错成本的关键是“换位思考”**：
三茂老师指出，真正“厉害”的活动从业者并非天生聪明，而是通过不断尝试、犯错、总结，将所有“坑”都走一遍，最终留下正确路径。因此，建议参与者以“换位思考”视角去理解用户需求与活动设计，从而减少盲目试错。
2. **试错是必经之路，但可被优化**：
会议强调，即使不参加听课，参与者也可以自行尝试（如“写5个点，不行就再写”），但通过“畅血卡”等报名机会，可获得专业指导与资源，显著降低试错成本。
3. **“活动大神”本质是“经历过所有错误的人”**：
三茂老师用“把所有错都走一遍”来解释高手的形成路径，打破“天赋论”，强调实践与反思的重要性。
4. **会议主题明确但未深入展开**：
虽然三茂老师提出“换位思考”作为核心，但未展开具体案例或操作步骤，也未深入讨论活动通过率、流程设计等细节，后续可进一步探讨。
**结论：**
本次会议以“换位思考”为核心理念，倡导参与者以用户视角出发，主动试错、反思、迭代，从而在活动策划中实现低成本、高效率的实践路径。参会者主要为对活动运营感兴趣的学习者或从业者，会议内容具有启发性但尚属初步引导。
---
输入 token 数: 360
输出 token 数: 487
kv cache 总长度: 8191
LLM 子图耗时统计:
g0: 7.223 ms
g1: 42.440 ms
g2: 72.761 ms
g3: 107.597 ms
g4: 152.216 ms
g5: 189.962 ms
g6: 223.133 ms
g7: 263.517 ms
g8: 301.006 ms
g9: 333.888 ms
g10: 362.304 ms
g11: 404.907 ms
g12: 442.639 ms
g13: 481.309 ms
g14: 520.008 ms
g15: 549.459 ms
g16: 588.102 ms
g17: 625.447 ms
LLM 后处理耗时:
qwen3_post.axmodel: 19.455 ms
最长 TTFT 耗时:
203804.475 ms
, 约
20
tokens/s
Deocde:
279.483 ms
, 约
3.5
tokens/s
参考：
技术讨论
- Github issues
- QQ 群: 139953715
Runs of AXERA-TECH 3D-Speaker-Meeting-Summary on huggingface.co
26
Total runs
-2
24-hour runs
-2
3-day runs
-2
7-day runs
4
30-day runs
More Information About 3D-Speaker-Meeting-Summary huggingface.co Model
Url of 3D-Speaker-Meeting-Summary
3D-Speaker-Meeting-Summary huggingface.co Url
Provider of 3D-Speaker-Meeting-Summary huggingface.co
Other API from AXERA-TECH
Updated:August 26 2026
Total runs:
213
Run Growth:
72
Growth Rate:
33.80%
Total runs:
200
Run Growth:
-311
Growth Rate:
-161.98%
Total runs:
95
Run Growth:
-366
Growth Rate:
-373.47%
Total runs:
62
Run Growth:
22
Growth Rate:
36.07%
Total runs:
57
Run Growth:
60
Growth Rate:
88.24%
Total runs:
57
Run Growth:
56
Growth Rate:
100.00%
Total runs:
47
Run Growth:
22
Growth Rate:
46.81%
Total runs:
47
Run Growth:
24
Growth Rate:
51.06%
Total runs:
45
Run Growth:
14
Growth Rate:
31.82%
Total runs:
45
Run Growth:
-1
Growth Rate:
-2.27%
Total runs:
43
Run Growth:
-1
Growth Rate:
-2.27%
Total runs:
43
Run Growth:
4
Growth Rate:
8.70%
Total runs:
42
Run Growth:
23
Growth Rate:
54.76%
Total runs:
42
Run Growth:
15
Growth Rate:
34.88%
Total runs:
41
Run Growth:
10
Growth Rate:
24.39%
Total runs:
41
Run Growth:
-197
Growth Rate:
-492.50%
Total runs:
39
Run Growth:
28
Growth Rate:
68.29%
Total runs:
37
Run Growth:
-9
Growth Rate:
-23.08%
Total runs:
34
Run Growth:
20
Growth Rate:
55.56%
Total runs:
34
Run Growth:
1
Growth Rate:
3.13%
Total runs:
34
Run Growth:
7
Growth Rate:
20.59%
Total runs:
33
Run Growth:
5
Growth Rate:
13.89%
Total runs:
33
Run Growth:
11
Growth Rate:
31.43%
Total runs:
33
Run Growth:
-1
Growth Rate:
-3.13%
Total runs:
32
Run Growth:
12
Growth Rate:
37.50%
Total runs:
32
Run Growth:
-39
Growth Rate:
-118.18%
Total runs:
32
Run Growth:
5
Growth Rate:
16.13%
Total runs:
31
Run Growth:
25
Growth Rate:
75.76%
Total runs:
31
Run Growth:
12
Growth Rate:
40.00%
Total runs:
31
Run Growth:
10
Growth Rate:
31.25%
Total runs:
31
Run Growth:
22
Growth Rate:
66.67%
Total runs:
31
Run Growth:
30
Growth Rate:
100.00%
Total runs:
31
Run Growth:
9
Growth Rate:
29.03%
Total runs:
31
Run Growth:
11
Growth Rate:
35.48%
Total runs:
30
Run Growth:
7
Growth Rate:
23.33%
Total runs:
30
Run Growth:
18
Growth Rate:
60.00%
Total runs:
28
Run Growth:
-16
Growth Rate:
-57.14%
Total runs:
28
Run Growth:
0
Growth Rate:
0.00%
Total runs:
28
Run Growth:
7
Growth Rate:
25.00%
Total runs:
28
Run Growth:
18
Growth Rate:
64.29%
Total runs:
27
Run Growth:
12
Growth Rate:
42.86%
Total runs:
26
Run Growth:
13
Growth Rate:
50.00%
Total runs:
26
Run Growth:
0
Growth Rate:
0.00%
Total runs:
25
Run Growth:
-34
Growth Rate:
-130.77%
Total runs:
24
Run Growth:
0
Growth Rate:
0.00%
Total runs:
23
Run Growth:
16
Growth Rate:
66.67%
Total runs:
23
Run Growth:
-2
Growth Rate:
-9.52%
Total runs:
22
Run Growth:
8
Growth Rate:
34.78%
Total runs:
22
Run Growth:
11
Growth Rate:
45.83%
Total runs:
21
Run Growth:
13
Growth Rate:
59.09%