This paper addresses the problem of noise estimation in the context of speech processing. A recently proposed quantile-based approach to noise estimation has the merit of not relying on the explicit detection of speech, non-speech boundaries. Here this approach is extended to both time and frequency. The resultant time-frequency quantile-based noise estimation is shown to give superior ASR performance. Results on the Aurora 2 Distributed Speech Recognition Database show an average relative performance improvement over the ETSI front-end baseline of 35%. The merits of the new system include: the relatively few parameters to optimise, the independence of absolute signal levels and minimal latency, all of which assist in real-time implementations.