Error report: Errors about run "BaiduCN1.2k Model" !

[gangyahaidao]

Hello guys, when i test Aishell, everything runs ok, except the high "Word Error Rate"!

then i begin to test "BaiduCN1.2k Model", something went wrong.

1.i known that BaiduCN1.2k speech rate is 8khz, then I change the hz from file demo_server.py and demo_client.py to 8000hz, when i speak from the mocrophone, i get some errors from demo_server.py, as follow:

Received utterance[length=94208] from 127.0.0.1, saved to demo_cache/20180419073019_127.0.0.1.wav.
/usr/local/lib/python2.7/dist-packages/resampy/core.py:90: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  if not np.issubdtype(x.dtype, np.float):

and when the warmup file is 8000hz, got the same error:

('Warm-up Test Case %d: %s', 0, u'/home/train/.cache/paddle/dataset/speech/Aishell/data_aishell/wav/test/S0913/BAC009S0913W0152.wav')
/usr/local/lib/python2.7/dist-packages/resampy/core.py:90: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  if not np.issubdtype(x.dtype, np.float):

I dont known why, please light me

2.question two: I dont known how to run asr server by "BaiduCN1.2k Model", here is my config file run_zh_demo_server.sh```

start demo server

CUDA_VISIBLE_DEVICES=0 \ python -u deploy/demo_server.py \ --host_ip='localhost' \ --host_port=8086 \ --num_conv_layers=2 \ --num_rnn_layers=3 \ --rnn_layer_size=1024 \ --alpha=1.15 \ --beta=0.15 \ --cutoff_prob=1.0 \ --cutoff_top_n=40 \ --use_gru=True \ --use_gpu=True \ --share_rnn_weights=False \ --speech_save_dir='demo_cache' \ --warmup_manifest='data/aishell/manifest.test' \ --mean_std_path='models/baidu_zh12k/mean_std.npz' \ --vocab_path='models/baidu_zh12k/vocab.txt' \ --model_path='models/baidu_zh12k/params.tar.gz' \ --lang_model_path='models/lm/zh_giga.no_cna_cmn.prune01244.klm' \ --decoding_method='ctc_beam_search' \ --specgram_type='linear'

after I startup the server, it cannot recognise wav file correct, what else I need to do?

any help will be very appreciate, thank you 

[frozenfires]

Question.2: You use gpu to run, but the parameter CUDA_VISIBLE_DEVICES=0. It should be set to the parameters of the language model. You can try 0, 1, 2, 4, 4. Or parameter use_gpu=False.

[gangyahaidao]

thank you for quick help, after i refer to #116 , i get this changeto my file "设置cutoff_prob小于1.0，比如0.99", amazy, error disappear! But, the recognise result is very bad, for example, i said "你好", i got this

Start Recording ...                                                               Speech[length=36864] Sent.
Recognition Results: 芯
Start Recording ...                                                                 Speech[length=38912] Sent.
Recognition Results: 芯
Start Recording ...                                                                                      Speech[length=49152] Sent.
Recognition Results: 匕

is this tell that the "BaiduCN1.2k Model" has some training problem, or it doesnot have trained enough well ??

[frozenfires]

set alpha=2.6,this is my config:

start demo server

CUDA_VISIBLE_DEVICES=0 \ python -u deploy/demo_server.py \ --host_ip='192.168.4.40' \ --host_port=6666 \ --beam_size=300 \ --num_conv_layers=2 \ --num_rnn_layers=3 \ --rnn_layer_size=1024 \ --alpha=2.6 \ --beta=5.0 \ --cutoff_prob=0.99 \ --cutoff_top_n=40 \ --use_gru=True \ --use_gpu=False \ --share_rnn_weights=False \ --speech_save_dir='demo_cache' \ --warmup_manifest='data/aishell/manifest.test' \ --mean_std_path='data/aishell/mean_std.npz' \ --vocab_path='data/aishell/vocab.txt' \ --model_path='data/aishell/params.tar.gz' \ --lang_model_path='data/aishell/zh_giga.no_cna_cmn.prune01244.klm' \ --decoding_method='ctc_beam_search' \ --specgram_type='linear'

[frozenfires]

I was very accurate when I tried, and I doubt your Chinese pronunciation. -_-!!

[gangyahaidao]

so kind of you ,but when i set alpha=2.6 and --beta=5.0 with "BaiduCN1.2k Model", I got this result, my chinese is study with a Beijing Mandarin frind, not bad i guess!, when i say "今天天气怎么样", got this:

Response Time: 3.740669, Transcript: 韦莱兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹
Received utterance[length=110592] from 127.0.0.1, saved to demo_cache/20180419084354_127.0.0.1.wav.
Response Time: 3.217678, Transcript: 韦莱兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹
Received utterance[length=114688] from 127.0.0.1, saved to demo_cache/20180419084403_127.0.0.1.wav.
Response Time: 3.348469, Transcript: 韦莱兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹韦兹

so .... Thank you anyway~ By the way, can i ask you what size of your model "baidu_cn1.2k_model.tar.gz", is it 759M ?? i doubt the file is complete downloaded

[frozenfires]

no,my speech model use the 'Mandarin Aishell Model 151h' Do you have QQ or WeChat? We can communicate and share easily.

[gangyahaidao]

I download it again, 759M, so my model file is complete, but why i get such strange output, much LOWER WER than Aishell, I will try to find out the reason

[gangyahaidao]

Yes, QQ 736791342 Wang, when i use --alpha=2.6 --beta=5.0 , its a little slower than --alpha=1.15 --beta=0.15, and donot have much higher accurate

[gangyahaidao]

this error solved, other error will create a new issue, close it

[blood0708]

@gangyahaidao 我也遇到了相同的问题，请问最后是怎么解决的阿？

[blood0708]

@gangyahaidao I also met the same problem. How did you solve it in the end?

[sunjunlishi]

@gangyahaidao how do you use it

[sunjunlishi]

@gangyahaidao how do you resolve it,very GOOD!!!!!!!!!!!!!!!!!!11

[code-R]

@gangyahaidao Can you please share the mandarin pretrained model.. the link available doesn't work any more. (BaiduCN1.2k Model)