Skip to content

Latest commit

 

History

History
119 lines (83 loc) · 7.65 KB

README.md

File metadata and controls

119 lines (83 loc) · 7.65 KB
Raw

wav2vec Unsupervised (wav2vec-U)

Wav2vec Unsupervised (wav2vec-U) and the 2.0 version are frameworks for building speech recognition systems without any labeled training data as described in Unsupervised Speech Recognition (Baevski et al., 2021) and Towards End-to-end Unsupervised Speech Recognition (Liu, et al., 2022). The model takes as input wav2vec 2.0 or XLSR representations (see pretrained models) as well as unlabeled speech and text data.

The training procedure consists of three consecutive main steps:

  • Preparation of speech representations and text data
  • Generative adversarial training (GAN)
  • Iterative self-training + Kaldi LM-decoding

Preparation of speech and text data

Similar to wav2vec 2.0, data folders contain {train,valid,test}.{tsv,wrd,phn} files, where audio paths are stored in tsv files, and word, letter or phoneme transcriptions are stored in .{wrd,ltr,phn}.

In /path/to/data/with_silence you need a train.tsv file as well as (optionally) {valid,test}.{tsv,wrd,phn}. It is nice to have 10h.{tsv,phn} files there too for reproducing the ablation study on layer selection. In /path/to/data/without_silence you have the same files, except .tsv files contain audios with silences removed using rVAD.

Pre-requisites:

  • set FAIRSEQ_ROOT environmental variable to your fairseq installation
  • set RVAD_ROOT environmental variable to a checkout of rVADfast
  • set KENLM_ROOT environmental variable to the location of KenLM binaries
  • install PyKaldi and set KALDI_ROOT environmental variable to the location of your kaldi installation. To use the version bundled with PyKaldi, you can use /path/to/pykaldi/tools/kaldi

Create new audio files without silences:

# create a manifest file for the set original of audio files
python $FAIRSEQ_ROOT/examples/wav2vec/wav2vec_manifest.py /dir/to/save/audio/files --ext wav --dest /path/to/new/train.tsv --valid-percent 0

python scripts/vads.py -r $RVAD_ROOT < /path/to/train.tsv > train.vads

python scripts/remove_silence.py --tsv /path/to/train.tsv --vads train.vads --out /dir/to/save/audio/files

python $FAIRSEQ_ROOT/examples/wav2vec/wav2vec_manifest.py /dir/to/save/audio/files --ext wav --dest /path/to/new/train.tsv --valid-percent 0.01

Next, we need to preprocess the audio data to better match phonemized text data:

# wav2vec-U
zsh scripts/prepare_audio.sh /dir/with/{train,test,valid}.tsv /output/dir /path/to/wav2vec2/model.pt 512 14
# wav2vec-U 2.0
zsh scripts/prepare_audio_v2.sh /dir/with/{train,test,valid}.tsv /output/dir /path/to/wav2vec2/model.pt 64 14

Note that if you have splits different than train/valid/test, you will need to modify this script. The thrid argument is the PCA dimensionality for wav2vec-U and the number of MFCC clusters for wav2vec-U 2.0. The last argument is the 0-based index of the layer from which to extract representations.

Now we need to prepare text data:

zsh scripts/prepare_text.sh language /path/to/text/file /output/dir 1000 espeak /path/to/fasttext/lid/model sil_prob

The fourth argument is minimum number observations of phones to keep. If your text corpus is small, you might want to reduce this number.

The fifth argument is which phonemizer to use. Supported values are espeak, espeak-ng, and G2P (english only).

Pre-trained fasttext LID models can be downloaded here.

The last argument is the probability to introduce silence (<SIL>) between the word boundaries. We found the value 0.25/0.5 works in general for wav2vec-U and the 2.0 version respectively, but you might want to vary for languages that are never tested.

Prepare TIMIT data

TIMIT transcripts include silence. Therefore VAD is not used for audio preprocessing, and we do not wrap transcripts with silences or insert random silence in between words.

To prepare TIMIT data for both the matched an unmatched setup:

bash scripts/prepare_timit.sh /dir/to/timit/raw/data /output/dir /path/to/wav2vec2/model.pt

Note that we assume the TIMIT distribution with capitalized directories and filenames are used (e.g., TRAIN/DR1/FCJF0/SA1.PHN).

Generative adversarial training (GAN)

We then use a GAN model to build a first unsupervised ASR model. The data preparation above of both speech features and text data is a necessary procedure that enables the generator to match speech to text in an unsupervised way.

Launching GAN training on top of preprocessed features, with default hyperparameters can be done with:

PREFIX=w2v_unsup_gan_xp

# For wav2vec-U, audio features are pre-segmented
CONFIG_NAME=w2vu
TASK_DATA=/path/to/features/precompute_unfiltered_pca512_cls128_mean_pooled

# For wav2vec-U 2.0, use raw audio features
CONFIG_NAME=w2vu2
TASK_DATA=/path/to/features/

# Unpaired text input
TEXT_DATA=/path/to/data/phones  # path to fairseq-preprocessed GAN data (phones dir)
KENLM_PATH=/path/to/data/phones/kenlm.phn.o4.bin  # KenLM 4-gram phoneme language model (LM data = GAN data here)

PYTHONPATH=$FAIRSEQ_ROOT PREFIX=$PREFIX fairseq-hydra-train \
    -m --config-dir config/gan \
    --config-name $CONFIG_NAME \
    task.data=${TASK_DATA} \
    task.text_data=${TEXT_DATA} \
    task.kenlm_path=${KENLM_PATH} \
    common.user_dir=${FAIRSEQ_ROOT}/examples/wav2vec/unsupervised \
    model.code_penalty=2,4 model.gradient_penalty=1.5,2.0 \
    model.smoothness_weight=0.5,0.75,1.0 'common.seed=range(0,5)'

Once we find the best checkpoint (chosen using unsupervised metric that combined language model perplexity and vocabulary usage), we can use it to generate phone labels (or word labels with an appropriate kaldi WFST):

python w2vu_generate.py --config-dir config/generate --config-name viterbi \
fairseq.common.user_dir=${FAIRSEQ_ROOT}/examples/wav2vec/unsupervised \
fairseq.task.data=/path/to/dir/with/features \
fairseq.common_eval.path=/path/to/gan/checkpoint \ 
fairseq.dataset.gen_subset=valid results_path=/where/to/save/transcriptions

The decoding without LM works best on the same adjacent-mean-pooled features that the gan was trained on, while decoding with LM works better on features before the adjacent timestep mean-pooling step (without the "_pooled" suffix).

While the generator of wav2vec-U 2.0 is trained with an output frequency of 16hz, we found decoding at a higher frequency produces better results. This can be done by adding decode_stride=1 or 2 to the argument.

Iterative self-training + Kaldi LM-decoding

After the GAN training provides a first unsupervised model, we can then progressively refine the quality of transcriptions using several iterations of semi-supervised learning. We perform two iterations: first, pseudo-label the training data with the unsupervised GAN model and train an HMM on the pseudo-labels. Second, we relabel the training data with the HMM and then fine-tune the original wav2vec 2.0 model using the HMM pseudo-labels with a CTC loss. Note that HMM models use phonemes as output, while wav2vec 2.0 use letter. Both are decoded using WFST decoders into words.

Please see this README for more instructions on how to do iterative self-training + Kaldi LM-decoding.

*** Note: these instructions are a work in progress and will be updated over the next few days