Remove dependence on speech_recognition for Wit

[qmac]

Although speech_recognition was initially very useful, its somewhat unnecessary now that the Wit API is the only converter using it. In addition, its stuck using a 2016 version of the API.

[tyarkoni]

Alternatively, we could leave it in and add a generic SpeechRecognitionAPIConverter class if that's possible given the speech_recognition architecture.

[qmac]

Hesitant to close this issue because I still don't love being stuck at the 2016 version of the API. I have a sneaking suspicion this is causing the occasional "Bad request" errors on Wit tests.

[qmac]

Closing for now, hopefully speech_recognition gets updated soon.