A pipeline for pre-processing speech data obtained from psycholinguistic experiments

Main steps:

1. Convert speech data into .wav files
2. Get transcriptions
3. Get timing measures
4. Manually check transcriptions and timing alignments
5. Export and analyze results

Open terminal.

Log into server computer:
ssh server address
e.g., ssh webexperiments@hjpatt-136.umd.edu

Type in password (nothing will appear when you type):
(type in server password)

Access the folder that contains your speech data
cd path/to/your_folder
e.g., cd Sites/Web_Experiments/Phillips/Rosa/uploads

Check that the folder contains your speech data
ls

Go back to the main folder
cd ..

Compress the folder containing speech files to zip:
zip -r name_of_zip ~/path/to/folder/containing/recording_files/
e.g., zip -r uploads.zip uploads/
→ compresses ‘uploads’ folder and name the zipped folder ‘uploads.zip’

Check that the zip file is created in the main folder:
ls

Exit the server:
exit

Download the zip file from the server to your local computer (requires server password):
scp -r server_address:~/path/to/zip_in_the_server ~/path/to/local_destination_folder
(type in server password)
e.g., scp -r webexperiments@hjpatt-136.umd.edu:Sites/Web_Experiments/Phillips/Rosa/recordings.zip ~/PlanetCloze/data/backup
→ downloads ‘recordings.zip’ to a local folder ‘backup’

Output of this section: a zip file containing speech files saved locally on your computer

Unzip all individual speech files using terminal:
cd path_to_folder
unzip \*.zip

Convert the .webm files into .wav files:
(This step requires having ffmpeg installed. You can install it using Homebrew, following the suggestion here: https://superuser.com/questions/624561/install-ffmpeg-on-os-x)
for i in *.webm; do ffmpeg -i "$i" "${i%.*}.wav"; done

Rename the original folder to “webm_wav”

Make a new folder “wav” and copy-paste only the .wav files into the folder

Output of Step 1: “wav” folder containing all speech files in .wav format

This step involves getting automatic speech-to-text transcriptions and speech timing alignments using APIs that are publicly available online.

For timing data, you can either use Chronset to get speech onset latencies, or MFA to get onset + timing measures for multi-word utterances.

Download the FastTranscribeAlign.zip file, unzip it, and upload the unzipped folder to your Google Drive directly under My Drive

Inside the uploaded FastTranscribeAlign folder:

Replace the empty wav folder with your local wav folder that contains all .wav speech files

Open FastTranscribeAlign.ipynb using Google Colaboratory and run the script following the instructions in the script

Output of Step 2: “tg_raw” folder containing textgrid files with transcriptions and timing data

Prepare tg_raw/tg_MFA_processed (depending on which option you chose for timing measures) and wav folders

Download the CheckAndEdit script from the Github repository

Open Praat

Praat --> Open Praat script --> CheckAndEdit --> Run

Designate paths:

• tg_path: path to tg_raw/tg_MFA_processed folder
• wav_path: path to wav folder

Check and edit transcriptions and alignments (changes are automatically saved)

Rename tg_raw/tg_MFA_processed to tg_checked

Upload tg_checked folder to the FastTranscribeAlign folder in Google Drive

Output of Step 3: “tg_checked” folder inside the “FastTranscribeAlign” folder in Google Drive; contains textgrid files with transcriptions + timing data that have been manually inspected

Open and run # generate results section in FastTranscribeAlign.ipynb

Output of Step 4: results.csv file containing transcriptions + timing measures from all textgrid files