MaixCAM MaixPy Keyword recognition

Introduction

MaixCAM has ported the Maix-Speech offline speech library, enabling continuous Chinese numeral recognition, keyword recognition, and large vocabulary speech recognition capabilities. It supports audio recognition in PCM and WAV formats, and can accept input recognition via the onboard microphone.

Maix-Speech

Maix-Speech is an offline speech recognition library specifically designed for embedded environments. It has been deeply optimized for speech recognition algorithms, significantly reducing memory usage while maintaining excellent recognition accuracy. For detailed information, please refer to the Maix-Speech Documentation.

Keyword recognition

from maix import app, nn

speech = nn.Speech("/root/models/am_3332_192_int8.mud")
speech.init(nn.SpeechDevice.DEVICE_MIC)

kw_tbl = ['xiao3 ai4 tong2 xue2',
          'ni3 hao3',
          'tian1 qi4 zen3 me yang4']
kw_gate = [0.1, 0.1, 0.1]

def callback(data:list[float], len: int):
    for i in range(len):
        print(f"\tkw{i}: {data[i]:.3f};", end=' ')
    print("\n")

speech.kws(kw_tbl, kw_gate, callback, True)

while not app.need_exit():
    frames = speech.run(1)
    if frames < 1:
        print("run out\n")
        break

Usage

1. Import the app and nn modules

from maix import app, nn

2. Load the acoustic model

speech = nn.Speech("/root/models/am_3332_192_int8.mud")

• You can also load the am_7332 acoustic model; larger models provide higher accuracy but consume more resources.

3. Choose the corresponding audio device

speech.init(nn.SpeechDevice.DEVICE_MIC)
speech.init(nn.SpeechDevice.DEVICE_MIC, "hw:0,0")   # Specify the audio input device

• This uses the onboard microphone and supports both WAV and PCM audio as input.

speech.init(nn.SpeechDevice.DEVICE_WAV, "path/audio.wav")   # Using WAV audio input

speech.init(nn.SpeechDevice.DEVICE_PCM, "path/audio.pcm")   # Using PCM audio input

• Note that WAV must be 16KHz sample rate with S16_LE storage format. You can use the arecord tool for conversion.

arecord -d 5 -r 16000 -c 1 -f S16_LE audio.wav

• When recognizing PCM/WAV , if you want to reset the data source, such as for the next WAV file recognition, you can use the speech.device method, which will automatically clear the cache:

speech.device(nn.SpeechDevice.DEVICE_WAV, "path/next.wav")

4. Set up the decoder

kw_tbl = ['xiao3 ai4 tong2 xue2',
          'ni3 hao3',
          'tian1 qi4 zen3 me yang4']
kw_gate = [0.1, 0.1, 0.1]

def callback(data:list[float], len: int):
    for i in range(len):
        print(f"\tkw{i}: {data[i]:.3f};", end=' ')
    print("\n")

speech.kws(kw_tbl, kw_gate, callback, True)

• The user can configure multiple decoders simultaneously. kws decoder is registered to output a list of probabilities for all registered keywords from the last frame. Users can observe the probability values and set their own thresholds for activation.

• When setting up the kws decoder, you need to provide a keyword list separated by spaces in Pinyin, a keyword probability threshold list arranged in order, and specify whether to enable automatic near-sound processing. If set to True, different tones of the same Pinyin will be treated as similar words to accumulate probabilities. Finally, you need to set a callback function to handle the decoded data.

• Users can also manually register near-sound words using the speech.similar method, with a maximum of 10 near-sound words registered for each Pinyin. (Note that using this interface to register near-sound words will override the near-sound table generated by enabling automatic near-sound processing.)

similar_char = ['zhen3', 'zheng3']
speech.similar('zen3', similar_char)

• If a decoder is no longer needed, you can deinitialize it by calling the speech.dec_deinit method.

speech.dec_deinit(nn.SpeechDecoder.DECODER_KWS)

5. Recognition

while not app.need_exit():
    frames = speech.run(1)
    if frames < 1:
        print("run out\n")
        break

• Use the speech.run method to run speech recognition. The parameter specifies the number of frames to run each time, returning the actual number of frames processed. Users can choose to run 1 frame each time and then perform other processing, or run continuously in a single thread, stopping it with an external thread.

• To clear the cache of recognized results, you can use the speech.clear method.

• When switching decoders during recognition, the first frame after the switch may produce incorrect results. You can use speech.skip_frames(1) to skip the first frame and ensure the accuracy of subsequent results.

Recognition Results

If the above program runs successfully, speaking into the onboard microphone will yield keyword recognition results, such as:

kws log 2.048s, len 24
decoder_kws_init get 3 kws
  00, xiao3 ai4 tong2 xue2
  01, ni3 hao3
  02, tian1 qi4 zen3 me yang4
find shared memory(491520),  saved:491520
    kw0: 0.959; 	kw1: 0.000; 	kw2: 0.000;     # xiao3 ai4 tong2 xue2
    kw0: 0.000; 	kw1: 0.930; 	kw2: 0.000;     # ni3 hao3
    kw0: 0.000; 	kw1: 0.000; 	kw2: 0.961;     # tian1 qi4 zen3 me yang4