I did some experimenting with Watson and Google using podcasts, and Watson came out pretty well in the mix - https://dague.net/2017/06/12/comparing-speech-recognition-fo... and https://dague.net/2017/06/13/visualizing-watson-speech-trans.... I haven't done systematic analysis with a lot of data points yet.

You can get per word confidence levels back from the Watson API if you ask for it, which is useful to know what parts it is sure about and what parts it is not.

Disclaimer: I work at IBM. But feel free to reach out at [email protected] if you want some help debugging on the Watson STT side.