Skip to content

Unsupervised phone and word segmentation using dynamic programming on self-supervised VQ features.


Notifications You must be signed in to change notification settings


Repository files navigation

Unsupervised Phone and Word Segmentation using Vector-Quantized Neural Networks

License: MIT


Unsupervised phone and word segmentation on speech data is performed. The experiments are described in:

  • H. Kamper, "Word segmentation on discovered phone units with dynamic programming and self-supervised scoring," IEEE/ACM Transactions on Audio, Speech and Language Processing, vol. 31, pp. 684-694, 2023. [arXiv]
  • H. Kamper and B. van Niekerk, "Towards unsupervised phone and word segmentation using self-supervised vector-quantized neural networks," in Proc. Interspeech, 2021. [arXiv]

Please cite these papers if you use the code.


Dependencies can be installed in a conda environment:

conda env create -f environment.yml
conda activate dpdp

This does not include wordseg, which should be installed in its own environment according to its documentation.

Install the DPDP AE-RNN package:

git clone ../dpdp_aernn

Minimal usage example: DPDP AE-RNN with DPDP CPC+K-means on Buckeye

In the sections that follow I give more complete details. In this section I briefly outline the sequence of steps that should reproduce the DPDP system results on Buckeye given in the paper. To apply the approach on other datasets you will need to carefully work through the subsequent sections, but I hope that this current section helps you to get going.

  1. Obtain the ground truth alignments for Buckeye provided in as part of this release. Extract it into data/. There should now be a data/buckeye/ directory with the alignments.

  2. Extract CPC+K-means features for Buckeye. Do this by following the steps in the CPC-big subsection below.

  3. Perform acoustic unit discovery using DPDP CPC+K-means:

    ./ --downsample_factor 1 --dur_weight 2 \
        --input_format=txt --algorithm=dp_penalized cpc_big buckeye val
  4. Perform word segmentation on the discovered units using the DPDP AE-RNN:

    ./ --algorithm=dpdp_aernn \
        cpc_big buckeye val phoneseg_dp_penalized
  5. Evaluate the segmentation:

    ./ cpc_big buckeye val \

The result should correspond approximately to the following on the Buckeye validation data:

Word boundaries:
Precision: 35.80%
Recall: 36.30%
F-score: 36.05%
OS: 1.40%
R-value: 45.13%
Word token boundaries:
Precision: 23.93%
Recall: 24.23%
F-score: 24.08%
OS: 1.24%

Example encodings: CPC-big features on Buckeye

Install the ZeroSpeech 2021 baseline system from my fork by following the steps in the installation section of the readme. Make sure that vqwordseg/ (this repository) and zerospeech2021_baseline/ are in the same directory.

From the vqwordseg/ directory, move to the ZeroSpeech 2021 directory:

cd ../zerospeech2021_baseline/

Extract individual Buckeye wav files:

./ ../datasets/buckeye/

The argument should point to your local copy of Buckeye.

Encode the Buckeye data:

conda activate zerospeech2021_baseline
./ wav/buckeye/val/ exp/buckeye/val/
./ wav/buckeye/test/ exp/buckeye/test/

conda activate dpdp
./ wav/buckeye/val/ exp/buckeye/hubert/val/
./ wav/buckeye/test/ exp/buckeye/hubert/test/

Move back and deactivate the environment:

cd ../vqwordseg/
conda activate dpdp

Dataset format and directory structure

This code should be usable with any dataset given that alignments and VQ encodings are provided.

For evaluation you need the ground truth phone and (optionally) word boundaries. These should be stored in the directories data/<dataset>/phone_intervals/ and data/<dataset>/word_intervals/ using the following filename format:


E.g., data/buckeye/phone_intervals/s01_01a_003222-003256.txt could consist of:

0 5 hh
5 10 iy
10 15 jh
15 19 ih
19 27 s
27 34 s
34 46 iy
46 54 m
54 65 z
65 69 l
69 78 ay
78 88 k

The duration-penalized dynamic programming (DPDP) algorithms operate on the output vector quantized (VQ) models. The (pre-)quantized representations and code indices should be provided in the exp/ directory. These are used as input to the VQ-segmentation algorithms; the segmented output is also produced in exp/.

As an example, the directory exp/vqcpc/buckeye/ should contain a file embedding.npy, which is the codebook matrix for a VQ-CPC model trained on Buckeye. This matrix will have the shape [n_codes, code_dim]. The directory exp/vqcpc/buckeye/val/ needs to contain at least subdirectories for the encoded validation set:

  • prequant/
  • indices/

The prequant/ directory contains the encodings from the VQ model before quantization. These encodings are given as text files with an embedding per line, e.g. the first three lines of prequant/s01_01a_003222-003256.txt could be:

 0.1601707935333252 -0.0403369292616844  0.4687763750553131 ...
 0.4489639401435852  1.3353070020675659  1.0353083610534668 ...
-1.0552909374237061  0.6382007002830505  4.5256714820861816 ...

The indices/ directory contains the code indices to which the auxiliary embeddings are actually mapped, i.e. which of the codes in embedding.npy are closest (under some metric) to the pre-quantized embedding. The code indices are again given as text files, with each index on a new line, e.g. the first three lines of indices/s01_01a_003222-003256.txt could be:


Any VQ model can be used. In the preceding section section I gave an example of using CPC-big with K-means; in the section below I give an example of how VQ-VAE and VQ-CPC can be used to obtain codes for the Buckeye dataset. In the subsequent section DPDP segmentation is described.

Example encodings: VQ-VAE and VQ-CPC on Buckeye

You can obtain the VQ input representations using the file format indicated above. As an example, here I describe how I did it for the Buckeye data.

First the following repositories need to be installed with their dependencies:

If you made sure that the dependencies are satisfied, these packages can be installed locally by running ./

Change directory to ../VectorQuantizedCPC and then perform the following steps there. Pre-process audio and extract log-Mel spectrograms:

./ in_dir=../datasets/buckeye/ dataset=buckeye

Encode the data and write it to the vqwordseg/exp/ directory. This should be performed for all splits (train, val and test):

./ checkpoint=checkpoints/cpc/english2019/ \
    split=val \
    save_indices=True \
    save_auxiliary=True \
    save_embedding=../vqwordseg/exp/vqcpc/buckeye/embedding.npy \
    out_dir=../vqwordseg/exp/vqcpc/buckeye/val/ \

Change directory to ../VectorQuantizedVAE and then run the following there. The audio can be pre-processed again (as above), or alternatively you can simply link to the audio from VectorQuantizedCPC/:

ln -s ../VectorQuantizedCPC/datasets/ .

Encode the data and write it to the vqwordseg/exp/ directory. This should be performed for all splits (train, val and test):

# Buckeye
./ checkpoint=checkpoints/2019english/ \
    split=train \
    save_indices=True \
    save_auxiliary=True \
    save_embedding=../vqwordseg/exp/vqvae/buckeye/embedding.npy \
    out_dir=../vqwordseg/exp/vqvae/buckeye/train/ \

You can delete all the created auxiliary_embedding1/ and codes/ directories since these are not used for segmentation.

Phone segmentation

DP penalized segmentation:

# Buckeye (GMM)
./ --downsample_factor 1 --input_format=npy \
    --algorithm=dp_penalized --dur_weight 0.001 \
    gmm buckeye val --output_tag phoneseg_merge

# Buckeye (VQ-CPC)
./ --input_format=txt --algorithm=dp_penalized \
    vqcpc buckeye val

# Buckeye (VQ-VAE)
./ vqvae buckeye val

# Buckeye (CPC-big)
./ --downsample_factor 1 --dur_weight 2 --input_format=txt \
    --algorithm=dp_penalized cpc_big buckeye val

# Buckeye (HuBERT)
./ --downsample_factor 2 --dur_weight 3 --input_format=npy \
    --algorithm=dp_penalized hubert buckeye val

# Buckeye (CPC-big) HSMM
./ --algorithm dp_penalized_hsmm --downsample_factor 1 \
    --dur_weight 1.0 --model_eos --dur_weight_func neg_log_gamma \
    --output_tag=phoneseg_hsmm_tune cpc_big buckeye val

# Buckeye Felix split (CPC-big) HSMM
./ --algorithm dp_penalized_hsmm --downsample_factor 1 \
    --dur_weight 1.0 --model_eos --dur_weight_func neg_log_gamma \
    --output_tag=phoneseg_hsmm_tune cpc_big buckeye_felix test

# Xitsonga (CPC-big)
./ --downsample_factor 1 --dur_weight 2 --input_format=txt \
    --algorithm=dp_penalized cpc_big xitsonga train

# Buckeye (XLSR)
./ --downsample_factor 2 --dur_weight 2500 \
    --input_format=npy --algorithm=dp_penalized xlsr buckeye val

# Buckeye (ResDAVEnet-VQ)
./ --downsample_factor 2 --dur_weight 3 --input_format=txt \
    --algorithm=dp_penalized resdavenet_vq buckeye val

# Buckeye (ResDAVEnet-VQ3)
./ --downsample_factor 4 --dur_weight 0.001 \
    --input_format=txt --algorithm=dp_penalized resdavenet_vq_quant3 \
    buckeye val --output_tag=phoneseg_merge

# Buckeye Felix split (VQ-VAE)
./ --output_tag=phoneseg_dp_penalized \
    vqvae buckeye_felix test

# Buckeye Felix split (CPC-big)
./ --downsample_factor 1 --dur_weight 2 \
    --output_tag=phoneseg_dp_penalized_tune cpc_big buckeye_felix val

# Buckeye Felix split (VQ-VAE) with Poisson duration prior
./ --output_tag=phoneseg_dp_penalized_poisson \
    --dur_weight_func neg_log_poisson --dur_weight 2 \
    vqvae buckeye_felix val

# Buckeye (VQ-VAE) with Gamma duration prior
./ --output_tag=phoneseg_dp_penalized_gamma \
    --dur_weight_func neg_log_gamma --dur_weight 15 vqvae buckeye val

# ZeroSpeech'17 English (CPC-big)
./ --downsample_factor 1 --dur_weight 2 --input_format=txt \
    --algorithm=dp_penalized cpc_big zs2017_en train

# ZeroSpeech'17 French (CPC-big)
./ --downsample_factor 1 --dur_weight 2 --input_format=txt \
    --algorithm=dp_penalized cpc_big zs2017_fr train

# ZeroSpeech'17 Mandarin (CPC-big)
./ --downsample_factor 1 --dur_weight 2 --input_format=txt \
    --algorithm=dp_penalized cpc_big zs2017_zh train

# ZeroSpeech'17 French (XLSR)
./ --downsample_factor 2 --dur_weight 1500 \
    --input_format=npy --algorithm=dp_penalized xlsr zs2017_fr train

# ZeroSpeech'17 Mandarin (XLSR)
./ --downsample_factor 2 --dur_weight 2500 \
    --input_format=npy --algorithm=dp_penalized xlsr zs2017_zh train

# ZeroSpeech'17 Lang2 (CPC-big)
./ --downsample_factor 1 --dur_weight 2 --input_format=txt \
    --algorithm=dp_penalized cpc_big zs2017_lang2 train

DP penalized N-seg. segmentation:

# Buckeye Felix split (VQ-VAE)
./ --algorithm=dp_penalized_n_seg \
    --n_frames_per_segment=3 --n_min_segments=3 vqvae buckeye_felix test

Evaluate segmentation:

# Buckeye (VQ-VAE)
./ vqvae buckeye val phoneseg_dp_penalized_n_seg

# Buckeye (CPC-big)
./ cpc_big buckeye val phoneseg_dp_penalized

Word segmentation

Word segmentation are performed on the segmented phone sequences.

Adaptor grammar word segmentation:

conda activate wordseg
# Buckeye (VQ-VAE)
./ --algorithm=ag vqvae buckeye val phoneseg_dp_penalized

# Buckeye (CPC-big)
./ --algorithm=ag cpc_big buckeye val phoneseg_dp_penalized

DPDP AE-RNN word segmentation:

# Buckeye (GMM)
./ --dur_weight=6 --algorithm=dpdp_aernn \
    gmm buckeye val phoneseg_dp_penalized

# Buckeye (CPC-big)
./ --algorithm=dpdp_aernn \
    cpc_big buckeye val phoneseg_dp_penalized

# Buckeye (HuBERT)
./ --dur_weight 6 --algorithm=dpdp_aernn \
    hubert buckeye val phoneseg_dp_penalized

DPDP AE-RNN word segmentation followed by K-means clustering on the AE-RNN embeddings:

# Buckeye (CPC-big)
./ --kmeans 14000 --algorithm=dpdp_aernn \
    cpc_big buckeye val phoneseg_dp_penalized

Perform K-means clustering on top of averaged features:

# Buckeye (HuBERT)
./ --kmeans 14000 hubert buckeye val wordseg_dpdp_aernn_dp_penalized

Evaluate the segmentation:

# Buckeye (VQ-VAE)
./ vqvae buckeye val wordseg_ag_dp_penalized

# Buckeye (CPC-big)
./ cpc_big buckeye val wordseg_ag_dp_penalized

# Buckeye (HuBERT)
./ hubert buckeye val wordseg_dpdp_aernn_dp_penalized

# Buckeye (HuBERT) with averaged embeddings
./ hubert buckeye val wordseg_dpdp_avgembed_kmeans14000

Evaluate the segmentation with the ZeroSpeech tools (deprecated, new version below):

./ cpc_big zs2017_zh train wordseg_segaernn_dp_penalized
cd ../zerospeech2017_eval/
ln -s \
    ~/endgame/projects/stellenbosch/vqseg/vqwordseg/exp/cpc_big/zs2017_zh/train/wordseg_dpdp_aernn_dp_penalized/clusters.txt \
conda activate zerospeech2020_updated
zerospeech2020-evaluate 2017-track2 . -l mandarin -o mandarin.json

Evaluate the segemntation with the updated ZeroSpeech tools:

cd ../zerospeech2017_eval/
ln -s \
    /media/kamperh/endgame/projects/stellenbosch/vqseg/vqwordseg/exp/hubert/zs2017_en/train/wordseg_dpdp_avgembed_kmeans43000/clusters.txt \
ln -s \
    /media/kamperh/endgame/projects/stellenbosch/vqseg/vqwordseg/exp/hubert/zs2017_fr/train/wordseg_dpdp_avgembed_kmeans29000/clusters.txt \
ln -s \
    /media/kamperh/endgame/projects/stellenbosch/vqseg/vqwordseg/exp/hubert/zs2017_zh/train/wordseg_dpdp_avgembed_kmeans3000/clusters.txt \
ln -s \
    /media/kamperh/endgame/projects/stellenbosch/vqseg/vqwordseg/exp/cpc_big/zs2017_lang1/train/wordseg_dpdp_aernn_dp_penalized/clusters.txt \
ln -s \
    /media/kamperh/endgame/projects/stellenbosch/vqseg/vqwordseg/exp/cpc_big/zs2017_lang2/train/wordseg_dpdp_aernn_dp_penalized/clusters.txt \
conda activate zrc-toolkit
zrc benchmarks:run tde17 2017_hubert


Print the word clusters:

./ cpc_big buckeye val wordseg_ag_dp_penalized

Listen to segmented codes:

./ vqvae buckeye val phoneseg_dp_penalized 343
./ vqvae buckeye val wordseg_tp_dp_penalized 486_
./ cpc_big buckeye val phoneseg_dp_penalized 50

This requires sox and that you change the path at the beginning of For ZeroSpeech'17 data, use instead.

Synthesize an utterance:

./ vqvae buckeye val phoneseg_dp_penalized \
cd ../VectorQuantizedVAE
./ checkpoints/2019english/ \
cd -

Complete example on ZeroSpeech data

An example of phone and word segmentation on the surprise language.

Encode data:

cd ../zerospeech2021_baseline
conda activate pytorch
./ path_to_data/datasets/zerospeech2020/2020/2017/ \
    zs2017_lang1 train

conda activate zerospeech2021_baseline
./ wav/zs2017_lang1/train/ exp/zs2017_lang1/train/

Phone segmentation:

cd ../vqwordseg
conda activate pytorch
# Create links in exp/cpc_big/
./ --downsample_factor 1 --dur_weight 2 --input_format=txt \
    --algorithm=dp_penalized cpc_big zs2017_lang1 train
./ cpc_big zs2017_lang1 train phoneseg_dp_penalized 3

Word segmentation:

./ --algorithm=dpdp_aernn cpc_big zs2017_lang1 train \
./ cpc_big zs2017_lang1 train \
    wordseg_dpdp_aernn_dp_penalized 33_10_11_14_1_34_

Convert to ZeroSpeech format:

./ cpc_big zs2017_lang1 train \

Evaluate the segmentation with the ZeroSpeech tools:

cd ../zerospeech2017_eval/
conda activate zerospeech2020
ln -s \
    ~/endgame/projects/stellenbosch/vqseg/vqwordseg/exp/cpc_big/zs2017_zh/train/wordseg_dpdp_aernn_dp_penalized/clusters.txt \
zerospeech2020-evaluate 2017-track2 . -l mandarin -o mandarin.json

About the Buckeye data splits

The particular split of Buckeye that I use in this repository is a legacy split with a somewhat complicated history. But in short the test set is exactly the same one used in the ZeroSpeech 2015 challenge. The remaining speakers were then used for a validation set and an additional held-out test set. This additional test set has the same number of speakers as the validation set, but most papers just report results on the ZeroSpeech 2105 test set.

The result is the following split of Buckeye, according to speaker:

  • Train (English1 in my thesis, devpart1 in other repos): s02, s03, s04, s05, s06, s08, s10, s11, s12, s13, s16, s38.
  • Validation (devpart2 in other repos): s17, s18, s19, s22, s34, s37, s39, s40.
  • Test (English2 in my thesis, ZS in other repos): s01, s20, s23, s24, s25, s26, s27, s29, s30, s31, s32, s33.
  • Additional test: s07, s09, s14, s15, s21, s28, s35, s36.

I fist used this in (Kamper et al., 2017) and since then in a number of follow-up papers. Others have also used this split, e.g. (Drexler and Glass, 2017), (Bhati et al., 2021), and ([Peng and Harwath, 2022]

Sets used in this repo. In this repo I only make use of the validation and test sets above, although features are extracted for the training set. See the experimental setup section of the paper.

The Kreuk split. Note that Kreuk et al. (2020) uses a different split which is also used by others. So in the section in the paper where I compare to their approach, I use their split:

  • Train: All Buckeye speakers not below.
  • Validation: s25, s36, s39, s40.
  • Test: s03, s07, s31, s34.

This split is not included in this repository---it made things too cluttered. And note that in the the paper I again don't use the Kreuk training set: I only report results on the test data when comparing to their models.

Reducing a codebook using clustering

If a codebook is very large, the codes could be reduced by clustering. The reduced codebook should be saved in a new model directory, and links to the original pre-quantized features should be created.

As an example, in cluster_codebook.ipynb, the ResDAVEnet-VQ codebook is loaded and reduced to 50 codes. The original codebook had 1024 codes, but only 498 of these were actually used; these are reduced to 50. The resulting codebook is saved to exp/resdavenet_vq_clust50/buckeye/embedding.npy. The pre-quantized features are linked to the original version in exp/resdavenet_vq/. The indices from the original model shouldn't be linked, since these doesn't match the new codebook (but an indices file isn't necessary for running many of the phone segmentation algorithms).

Number of clusters

For Buckeye experiments, we use 14000 classes, based on:


This is roughly the same number of clusters if you combine all the individual speaker-dependent clusters from the systems used in the Buckeye experiments in (Kamper et al., 2017).

For the ZeroSpeech experiments, we use the following number of clusters:

  • English: 43000 (42567)
  • French: 29000 (28919)
  • Mandarin: 3000 (2986)
  • Lang1 (German): 29000 (28954)
  • Lang2 (Wolof): 3500 (3592)

This is based on:


which matches the ZeroSpeech experiments from (Kamper et al., 2017).

Old work-flow (deprecated)

  1. Extract CPC+K-means features in ../zerospeech2021_baseline/.
  2. Perform phone segmentation here using
  3. Move to ../seg_aernn/notebooks/ and perform word segmentation.
  4. Move back here and evaluate the segmentation using
  5. For ZeroSpeech systems, the evaluation is done in ../zerospeech2017_eval/.


The code provided here is not pretty. But research should be reproducible. I provide no guarantees with the code, but please let me know if you have any problems, find bugs or have general comments.


Unsupervised phone and word segmentation using dynamic programming on self-supervised VQ features.