Welcome to sherpa-ncnn Discussions!

[csukuangfj]

👋 Welcome!

We’re using Discussions as a place to connect with other members of our community. We hope that you:

• Ask questions you’re wondering about.
• Share ideas.
• Engage with other community members.
• Welcome others and are open-minded. Remember that this is a community we
  build together 💪.

To get started, comment below with an introduction of yourself and tell us about what you do with this community.

[hieudevsoft]

I want implement into my project ? what should i do?

[hieudevsoft]

yes i have follow and completelly tutorial and workfine but i copy assets and jniLibs to my own project then it not working :(

[hieudevsoft]

java.lang.UnsatisfiedLinkError: No implementation found for long hieubui.eup.androidspeech.SherpaNcnn.newFromAsset(android.content.res.AssetManager, hieubui.eup.androidspeech.RecognizerConfig) (tried Java_hieubui_eup_androidspeech_SherpaNcnn_newFromAsset and Java_hieubui_eup_androidspeech_SherpaNcnn_newFromAsset__Landroid_content_res_AssetManager_2Lhieubui_eup_androidspeech_RecognizerConfig_2)
it's log

[csukuangfj]

please post the complete logs related to sherpa-ncnn. You can use ncnn as the filer in the logfiler.

[csukuangfj]

Also, please first learn the sherpa-ncnn android project and learn how sherpa-ncnn is used in Android.

[csukuangfj]

https://github.com/k2-fsa/sherpa-ncnn/blob/master/.github/scripts/test-jni.sh

This script shows how to use sherpa-ncnn in kotlin and it is not related to android.

[temcgraw]

Words are detected in silence. In microphone example, if the samples provided to s->AcceptWaveform(...) are all zero then a series of words are continuously detected. This doesn't seem to happen in the unmodified example, but if I use another audio library (miniaudio) the noise is lower and samples are closer to zero in silence.

[temcgraw]

To see the issue make the RecordCallback be:
std::vector test(frames_per_buffer, 0.0f);
s->AcceptWaveform(16000, reinterpret_cast<const float *>(test.data()), frames_per_buffer);
return paContinue;

[ailoli1314]

[csukuangfj]

Sorry, the code is hard to read. Could you reformat it or could you upload your python script as an attachment?

[ailoli1314]

thanks，i had see the help for Endpointing,now i know how to use this

[acjf]

我在很早的版本体验还不错。但用最近的代码体验很差，相差真的太大了。以前2022-12-17-mixed-english-chinese-sherpa-ncnn.apk的代码版本还能支持吗，我很喜欢这个的体验

[csukuangfj]

信息有限，无法提供帮助

[acjf]

https://huggingface.co/csukuangfj/sherpa-ncnn-apk/blob/main/2022-12-17-mixed-english-chinese-sherpa-ncnn.apk我想要产出这个apk的这套代码。，在这个之后的都不行，体验都很差，只有这个的很流畅

[csukuangfj]

信息有限，无法提供帮助

什么叫不行？什么叫体验很差？能否具体描述？

[acjf]

1、下载后面的版本进行编译so,这步是正常能编译的
2、进行android示例，把sherpa-ncnn-conv-emformer-transducer-2022-12-06的模型放进去
3、运行示例，基本都是会发生崩溃，无法使用

[csukuangfj]

最后一遍了, 信息有限，无法提供帮助

[acjf]

1、下载2.1.15源码
2、执行build-android-arm64-v8a.sh,得到so，拷贝android 示例的arm64-v8a目录
3、拷贝模型放到assets文件夹
4、运行，闪退：日志：
Accessing hidden method Landroid/view/View;->computeFitSystemWindows(Landroid/graphics/Rect;Landroid/graphics/Rect;)Z (greylist, reflection, allowed)
Accessing hidden method Landroid/view/ViewGroup;->makeOptionalFitsSystemWindows()V (greylist, reflection, allowed)
Start to initialize model
------config------
RecognizerConfig(feat_config=FeatureExtractorConfig(sampling_rate=16000, feature_dim=80), model_config=ModelConfig(encoder_param="sherpa-ncnn-conv-emformer-transducer-2022-12-06/encoder_jit_trace-pnnx.ncnn.int8.param", encoder_bin="sherpa-ncnn-conv-emformer-transducer-2022-12-06/encoder_jit_trace-pnnx.ncnn.int8.bin", decoder_param="sherpa-ncnn-conv-emformer-transducer-2022-12-06/decoder_jit_trace-pnnx.ncnn.param", decoder_bin="sherpa-ncnn-conv-emformer-transducer-2022-12-06/decoder_jit_trace-pnnx.ncnn.bin", joiner_param="sherpa-ncnn-conv-emformer-transducer-2022-12-06/joiner_jit_trace-pnnx.ncnn.int8.param", joiner_bin="sherpa-ncnn-conv-emformer-transducer-2022-12-06/joiner_jit_trace-pnnx.ncnn.int8.bin", tokens="sherpa-ncnn-conv-emformer-transducer-2022-12-06/tokens.txt", encoder num_threads=1, decoder num_threads=1, joiner num_threads=1), decoder_config=DecoderConfig(method="greedy_search", num_active_paths=4), endpoint_config=EndpointConfig(rule1=EndpointRule(must_contain_nonsilence=False
RecognizerConfig(feat_config=FeatureExtractorConfig(sampling_rate=16000, feature_dim=80), model_config=ModelConfig(encoder_param="sherpa-ncnn-conv-emformer-transducer-2022-12-06/encoder_jit_trace-pnnx.ncnn.int8.param", encoder_bin="sherpa-ncnn-conv-emformer-transducer-2022-12-06/encoder_jit_trace-pnnx.ncnn.int8.bin", decoder_param="sherpa-ncnn-conv-emformer-transducer-2022-12-06/decoder_jit_trace-pnnx.ncnn.param", decoder_bin="sherpa-ncnn-conv-emformer-transducer-2022-12-06/decoder_jit_trace-pnnx.ncnn.bin", joiner_param="sherpa-ncnn-conv-emformer-transducer-2022-12-06/joiner_jit_trace-pnnx.ncnn.int8.param", joiner_bin="sherpa-ncnn-conv-emformer-transducer-2022-12-06/joiner_jit_trace-pnnx.ncnn.int8.bin", tokens="sherpa-ncnn-conv-emformer-transducer-2022-12-06/tokens.txt", encoder num_threads=1, decoder num_threads=1, joiner num_threads=1), decoder_config=DecoderConfig(method="greedy_search", num_active_paths=4), endpoint_config=EndpointConfig(rule1=EndpointRule(must_contain_nonsilence=False, min_trailing_sile
AAssetManager_open sherpa-ncnn-conv-emformer-transducer-2022-12-06/encoder_jit_trace-pnnx.ncnn.int8.param failed
Failed to load sherpa-ncnn-conv-emformer-transducer-2022-12-06/encoder_jit_trace-pnnx.ncnn.int8.param
Fatal signal 11 (SIGSEGV), code 1 (SEGV_MAPERR), fault addr 0x0 in tid 13896 (fsa.sherpa.ncnn), pid 13896 (fsa.sherpa.ncnn)
pid: 13896, tid: 13896, name: fsa.sherpa.ncnn >>> com.k2fsa.sherpa.ncnn <<<
---------------------------- PROCESS STARTED (13928) for package com.k2fsa.sherpa.ncnn ----------------------------
---------------------------- PROCESS ENDED (13885) for package com.k2fsa.sherpa.ncnn ----------------------------
#00 pc 0000000000077888 /data/app/~~PrGNfLChBdC4ICk_-eIt6Q==/com.k2fsa.sherpa.ncnn-wvVapGhciYYYbRlJz99jsQ==/lib/arm64/libsherpa-ncnn-jni.so (BuildId: f08eec7c2f0f2a56d7576ad2d3a166718938bb8e)
#1 pc 000000000005b8f4 /data/app/~~PrGNfLChBdC4ICk_-eIt6Q==/com.k2fsa.sherpa.ncnn-wvVapGhciYYYbRlJz99jsQ==/lib/arm64/libsherpa-ncnn-jni.so (BuildId: f08eec7c2f0f2a56d7576ad2d3a166718938bb8e)
#2 pc 000000000005a104 /data/app/~~PrGNfLChBdC4ICk_-eIt6Q==/com.k2fsa.sherpa.ncnn-wvVapGhciYYYbRlJz99jsQ==/lib/arm64/libsherpa-ncnn-jni.so (Java_com_k2fsa_sherpa_ncnn_SherpaNcnn_newFromAsset+336) (BuildId: f08eec7c2f0f2a56d7576ad2d3a166718938bb8e)
#10 pc 0000000000003a02 [anon:dalvik-classes3.dex extracted in memory from /data/app/~~PrGNfLChBdC4ICk_-eIt6Q==/com.k2fsa.sherpa.ncnn-wvVapGhciYYYbRlJz99jsQ==/base.apk!classes3.dex] (com.k2fsa.sherpa.ncnn.SherpaNcnn.+30)
#13 pc 0000000000001cda [anon:dalvik-classes3.dex extracted in memory from /data/app/~~PrGNfLChBdC4ICk_-eIt6Q==/com.k2fsa.sherpa.ncnn-wvVapGhciYYYbRlJz99jsQ==/base.apk!classes3.dex] (com.k2fsa.sherpa.ncnn.MainActivity.initModel+130)
#16 pc 0000000000001d4e [anon:dalvik-classes3.dex extracted in memory from /data/app/~~PrGNfLChBdC4ICk_-eIt6Q==/com.k2fsa.sherpa.ncnn-wvVapGhciYYYbRlJz99jsQ==/base.apk!classes3.dex] (com.k2fsa.sherpa.ncnn.MainActivity.onCreate+50)
---------------------------- PROCESS ENDED (13896) for package com.k2fsa.sherpa.ncnn ----------------------------

[acjf]

每一个tag的版本，我均尝试过，都是不同的表现，这里仅提供了最新版本的崩溃日志

[acjf]

模型使用的是：https://k2-fsa.github.io/sherpa/ncnn/pretrained_models/conv-emformer-transducer-models.html#csukuangfj-sherpa-ncnn-conv-emformer-transducer-2022-12-06-chinese-english 下载的模型

[csukuangfj]

感谢分享详细的 log! 我们现在可以提供帮助了。

AAssetManager_open sherpa-ncnn-conv-emformer-transducer-2022-12-06/encoder_jit_trace-pnnx.ncnn.int8.param failed
Failed to load sherpa-ncnn-conv-emformer-transducer-2022-12-06/encoder_jit_trace-pnnx.ncnn.int8.param

这个是关键的 log

请把你 assets 文件夹，在 android stduio 里，截个图，看里面有什么。

[csukuangfj]

我看到问题，我试修改下

太好啦！你终于会自己调试了

[acjf]

很抱歉，是我的问题，我犯了一个低级的错误。感谢你的耐心

[csukuangfj]

如果你觉得有用，请给我们点 star, 感谢支持!

[acjf]

咱们的推理框架支持耳语的模型吗，或者说耳语的能转成ncnn所需的模型吗

[csukuangfj]

如前所述，信息有限，无法回答。

[csukuangfj]

友情提示，如果你用最新的代码，只需要 libncnn.so 和 libsherpa-ncnn-jni.so, 其他的 .so 和 .a 文件，都可以删掉。

[sys-dotcom]

我现在有个需求，是做一个本地化部署的中英文语音翻译，就是讲话时，可以实时出现对照的中英文，并且可以降噪，不知道这个是否可以支持？如果是可以的，可以进行合作

[sys-dotcom]

目前，我这有大量可以支撑训练的语料，之前用了科大讯飞的，但是效果不理想，识别率不足90%，因为语料很多是船舶行业的专有名词，所以，这个语言模型是否可以支持训练？