Implementation of "SpeechLMScore: Evaluating speech generation using speech langauge model" https://arxiv.org/abs/2212.04559
You can install required python packages as:
python setup.py install
Download these pretrained models and update their path in run.sh.
Note: tokens.txt is located with speech ulm model.
Run the following command to download all the above models:
./download_pretrained_models.shGenerates speechlmscore for each file in audio_dir in file ppl.
Audio files with sampling rate of 16kHZ are supported.
Note: for using audio files other than .wav set ext variable is run.sh.
audio_dir=<folder containing audio>
layer=<Hubert layer to extract features>
./run.sh ${audio_dir} ${layer}
Additionally speech language model can be trained and used for evaluation as well. More Details