Aims: We present new extraction and identification techniques for supernova (SN) spectra developed within the Supernova Legacy Survey (SNLS) collaboration. Methods: The new spectral extraction method takes full advantage of photometric information from the Canada-France-Hawaï telescope (CFHT) discovery and reference images by tracing the exact position of the supernova and the host signals on the spectrogram. When present, the host spatial profile is measured on deep multi-band reference images and is used to model the host contribution to the full (supernova + host) signal. The supernova is modelled as a Gaussian function of width equal to the seeing. A χ2 minimisation provides the flux of each component in each pixel of the 2D spectrogram. For a host-supernova separation greater than ⪆1 pixel, the two components are recovered separately and we do not use a spectral template in contrast to more standard analyses. This new procedure permits a clean extraction of the supernova separately from the host in about 70% of the 3rd year ESO/VLT spectra of the SNLS. A new supernova identification method is also proposed. It uses the SALT2 spectrophotometric template to combine the photometric and spectral data. A galaxy template is allowed for spectra for which a separate extraction of the supernova and the host was not possible. Results: These new techniques have been tested against more standard extraction and identification procedures. They permit a secure type and redshift determination in about 80% of cases. The present paper illustrates their performances on a few sample spectra. Based on observations obtained with FORS1 at the Very Large Telescope on the Cerro Paranal, operated by the European Southern Observatory, Chile (ESO Large Programmes 171.A-0486 and 176.A-0589).