Code for "Homophone Reveals the Truth: A Reality Check for Speech2Vec"
A brief version of this report has been accepted by ICASSP-2023: "A Reality Check and A Practical Baseline for Semantic Speech Embedding"

• Free GPU RAM >= 3GB
• Free System RAM >= 30GB
• Pytorch Version >= 1.8.2

Download the dataset.zip (2.4GB) from Google Drive. Unzip it to the project root. Its structure is shown below:

dataset/
├── info
│   ├── 500h_word2wav_keys.pkl
│   ├── 500h_word_counter.pkl
│   ├── 500h_word_split.pkl
│   └── eval
│       ├── all_words_5846.pkl
│       ├── files
│       │   ├── EN-MC-30.txt
│       │   ├── EN-MEN-TR-3k.txt
│       │   ├── EN-MTurk-287.txt
│       │   ├── EN-MTurk-771.txt
│       │   ├── EN-RG-65.txt
│       │   ├── EN-RW-STANFORD.txt
│       │   ├── EN-SIMLEX-999.txt
│       │   ├── EN-SimVerb-3500.txt
│       │   ├── EN-VERB-143.txt
│       │   ├── EN-WS-353-ALL.txt
│       │   ├── EN-WS-353-REL.txt
│       │   ├── EN-WS-353-SIM.txt
│       │   └── EN-YP-130.txt
│       └── homophone.txt
├── split_mfcc_dict.pkl
└── split_mfcc_mean_std.pkl

We also provided some speech sentence segment exmaples in SentenceSegmentExamples.zip.

This instruction describes how we generated those files.

Just run python 1_train.py

The full training process (500-epoch) takes 8.4 days on an AMD 3900XT + RTX3090 machine.

use wandb to view the training process:

1. Create .wb_config.json file in the project root, using the following content:

   {
     "WB_KEY": "Your wandb auth key"
   }
   

2. add --dryrun=False to the training command, for example: python 1_train.py --dryrun=False

The checkpoints and embeddings of every epoch are in Full500EpochModelsEmbedings.zip (2.9GB)

The Rand Init model corresponds to: epoch-01_ws0.10_men0.08_loss-1.000000.pkl

The 500-Epoch model corresponds to: epoch499_ws0.15_men0.08_loss0.247943.pkl

Feel free to contact me if you have any question:
Email: [email protected]
WeChat: