Speechrecognition

Small bugfix release:

* Handle case when GSR doesn't return a confidence value (thanks jcsilva!).
* Config, style, and release improvements.
* Fix console window sometimes popping up when on Windows (thanks Qdrew!)
* Switch release over to universal Wheels rather than source distribution.

This is more of a maintenance release, but a few features slipped in as well:
- **Support for the Google Cloud Speech API** with `recognizer_instance.recognize_google_cloud` (thanks Thynix!), plus documentation and examples.
- **Automatic sample rate detection** in `speech_recognition.Microphone` - this should fully resolve all the "Invalid sample rate" issues from PyAudio.
- Project now has **automated tests and continuous integration** with TravisCI. It's pretty nifty, and has already caught a few things during development!
- Keywords example for `recognizer_instance.recognize_sphinx`.
- Documentation improvements and updated advice in troubleshooting and library reference.
- Bugfix - Google Speech Recognition sometimes didn't return the text with the highest confidence (thanks akabraham!).
- Bugfix - `EOFError` upon encountering malformed audio files; a proper exception message is now given.
- Updated FLAC binaries for OS X.
- Bugfix - invalid FLAC binary path on OS X (thanks akabraham!).
- Code cleanup.

- **Support for the Houndify API** with `recognizer_instance.recognize_houndify` (thanks tb0hdan!).
- `recognize_sphinx` now supports **keyword-based matching** via the `keywords=[("cat", 30), ("potato", 45)]` parameter.
- The second number in each pair is the sensitivity, which determines how loosely Sphinx will interpret speech to be those keywords - higher numbers mean more false positives, while lower numbers mean a lower detection rate.
- A new example for keyword matching is now available.
- **BREAKING CHANGE: API.AI STT API IS BEING SHUT DOWN SOON.** ([source](https://docs.api.ai/docs/query))
- For now, the `recognize_api` function will keep working if you're on a paid API.AI plan, and we will not be removing it until the service is shut down entirely.
- It is best to transition to another backend as soon as possible. I recommend Microsoft Bing Voice Recognition or Wit.ai for previous API.AI users.
- `phrase_time_limit` option for listening functions, to limit phrase lengths to a certain number of seconds.
- Support for operation timeouts with `recognizer_instance.operation_timeout` - this can be used to ensure long requests always take finite time.
- `recognize_ibm` now opts out of request logging by default, for improved user privacy (thanks michellemorales!). **This is a breaking change if you previously relied on request logging behaviour**.
- Bugfix - `listen()` sometimes didn't terminate on finite-length streams.
- Bugfix - Microsoft Bing Voice Recognition changed their authentication API endpoint, so that required some small code updates (thanks tmator!).
- Bugfix - 24-bit audio now works correctly on Python 2.
- Update Wit.ai API version from deprecated version.
- A bunch of documentation updates, fixes, and improvements.

Bugfix release.

Changes:
- api.ai now requires the `sessionId` field, so we'll just add that in (thanks jhoelzl!).
- Improve documentation a bit.
- Various other small fixes.

Changes:
- Bug fix: non-24-bit audio wasn't converted properly to 16-bit audio on Python 2, due to the new 24-bit audio shim. Thanks to jhoelzl for reporting!

Maintenance release:
- Python versions less than 3.4 don't support 24-bit audio properly. We now have pure-Python shims that will **allow 24-bit audio to work on those old Python versions**, though they will be somewhat slower. Thanks to danse for reporting the issue!
- Added **updated Pocketsphinx binaries and Pocketsphinx installation procedures** to match improvements on their end.
- Fix Unicode file paths on Windows.
- Fix caching in `recognizer_instance.recognize_bing`.
- We now use the Manylinux Docker image for building FLAC. Hopefully, this will make building universal Linux binaries easier for packagers.