如何在wit.ai中发送分块的音频数据进行语音识别？

Question

I have a large mp3 file(about 1.8GB), which I have to transcribe using wit.ai. Since I am working with wav files a lot, i converted it to wav file.

But since wit.ai's speech api can't take more than 10s long audio, I am planning to stream the file in chunks. But some how I am only getting reponse 400(bad request). I am not able to find out, what I am sending wrong. Following are the details:

headers = {'authorization': 'Bearer ' + wit_access_token,
         'Content-Type': 'audio/wav','Transfer-encoding': 'chunked'}
with open('meeting-record.wav', 'rb') as f:
    audio = f.read(2048)  # taken it any number
resp = requests.post(API_ENDPOINT, headers = headers,
                 data = audio)
print(resp) 
data = json.loads(resp.content)
text = data['_text']
print(text)
f.close()

I am getting the following output

<Response [400]>
Traceback (most recent call last):
  File ".\sound-record.py", line 61, in <module>
    text = data['_text']
KeyError: '_text'

Can someone show some pointers, where its going wrong?

Answer 1

I haven't used the wit.ai API before, but the Bing Speech API appears to require the data in a similar fashion. I'm not sure if you were getting the error because of your code, but in order to properly chunk and stream the file, you could add another function in there like this:

def stream_audio_file(speech_file, chunk_size=1024):
    # Chunk audio file
    with open(speech_file, 'rb') as f:
        while 1:
            data = f.read(1024)
        if not data:
            break
        yield data

Now as long as you have that function somewhere in your file to stream and chunk the data for you, you can go back to your initial method:

headers = {
    'Accept': 'application/json',
    'Transfer-Encoding': 'chunked',
    'Content-type': 'audio/wav',
    'Authorization': 'Bearer {0}'.format(YOUR_AUTH_TOKEN)
}

data = stream_audio_file(YOUR_AUDIO_FILE)

r = requests.post(url, headers=headers, data=data)

results = json.loads(r.content)

print(results)

Side Note: You mentioned you wanted something on your own server. There's a nice module called pocketsphinx , which is free, hosted on your machine, and written in Python. It pairs really well with the SpeechRecognition module, which provides a decent layer on top so you don't have to spend as much time formatting your requests.

Answer 2

Wit.ai is not supposed to transcribe long files, it is a system for recognizing short commands. You'd better use proper services:

• Google speech API

• Speechmatics

And many others