Well - I haven't tried the Speech Service myself . But as dev, we notices pretty high error rates with the 'user modeling service' as well.