readtextgrid parses Praat textgrids into tidy R dataframes.
Features
-
Simple: Minimal package with two core functions (
read_textgrid()andread_textgrid_lines()). - Tidy: Returns rectangular tibbles ready for downstream processing with dplyr and tidyr.
- Flexible: Supports both long and short textgrid file formats.
- Fast: Uses a compiled C++ tokenizer for high-throughput parsing.
Installation
Install readtextgrid from CRAN:
install.packages("readtextgrid")Development version. Install precompiled version of readtextgrid from R-universe:
install.packages(
"readtextgrid",
repos = c("https://tjmahr.r-universe.dev", "https://cloud.r-project.org")
)Basic usage
Here is the example textgrid created by Praat. It was created using New > Create TextGrid... with default settings in Praat.
This textgrid is bundled with this R package. We can locate the file with example_textgrid(). We read in the textgrid with read_textgrid().
library(readtextgrid)
# Locates path to an example textgrid bundled with this package
tg <- example_textgrid()
read_textgrid(path = tg)
#> # A tibble: 3 × 10
#> file tier_num tier_name tier_type tier_xmin tier_xmax
#> <chr> <int> <chr> <chr> <dbl> <dbl>
#> 1 Mary_John_bell.TextGrid 1 Mary IntervalTier 0 1
#> 2 Mary_John_bell.TextGrid 2 John IntervalTier 0 1
#> 3 Mary_John_bell.TextGrid 3 bell TextTier 0 1
#> xmin xmax text annotation_num
#> <dbl> <dbl> <chr> <int>
#> 1 0 1 "" 1
#> 2 0 1 "" 1
#> 3 NA NA <NA> NAThe dataframe contains one row per annotation: one row for each interval on an interval tier and one row for each point on a point tier. If a point tier has no points, it is represented with single row with NA values.
The columns encode the following information:
-
filefilename of the textgrid. By default this column uses the filename inpath. A user can override this value by setting thefileargument inread_textgrid(path, file), which can be useful if textgrids are stored in speaker-specific folders. -
tier_numthe number of the tier (as in the left margin of Praat’s textgrid editor) -
tier_namethe name of the tier (as in the right margin of Praat’s textgrid editor) -
tier_typethe type of the tier."IntervalTier"for interval tiers and"TextTier"for point tiers (this is the terminology used inside of the textgrid file format). -
tier_xmin,tier_xmaxstart and end times of the tier in seconds -
xmin,xmaxstart and end times of the textgrid interval or point tier annotation in seconds -
textthe text in the annotation -
annotation_numthe number of the annotation in that tier (1 for the first annotation, etc.)
Reading in directories of textgrids
Suppose we have data on multiple speakers with one folder of textgrids per speaker. As an example, this package has a folder called speaker_data bundled with it representing 5 five textgrids from 2 speakers.
📂 speaker-data
├── 📂 speaker001
│ ├── s2T01.TextGrid
│ ├── s2T02.TextGrid
│ ├── s2T03.TextGrid
│ ├── s2T04.TextGrid
│ └── s2T05.TextGrid
└── 📂 speaker002
├── s2T01.TextGrid
├── s2T02.TextGrid
├── s2T03.TextGrid
├── s2T04.TextGrid
└── s2T05.TextGridFirst, we create a vector of file-paths to read into R.
# Get the path of the folder bundled with the package
data_dir <- system.file(package = "readtextgrid", "speaker-data")
# Get the full paths to all the textgrids
paths <- list.files(
path = data_dir,
pattern = "TextGrid$",
full.names = TRUE,
recursive = TRUE
)We can use purrr::map()–map the read_textgrid() function over the paths—to read all these textgrids into R and combine them from a list to a single dataframe with purrr::list_rbind(). But note that this way doesn’t track any speaker information.
library(purrr)
paths |>
map(read_textgrid) |>
list_rbind()
#> # A tibble: 150 × 10
#> file tier_num tier_name tier_type tier_xmin tier_xmax xmin
#> <chr> <int> <chr> <chr> <dbl> <dbl> <dbl>
#> 1 s2T01.TextGrid 1 words IntervalTier 0 1.35 0
#> 2 s2T01.TextGrid 1 words IntervalTier 0 1.35 0.297
#> 3 s2T01.TextGrid 1 words IntervalTier 0 1.35 0.522
#> 4 s2T01.TextGrid 1 words IntervalTier 0 1.35 0.972
#> 5 s2T01.TextGrid 2 phones IntervalTier 0 1.35 0
#> 6 s2T01.TextGrid 2 phones IntervalTier 0 1.35 0.297
#> 7 s2T01.TextGrid 2 phones IntervalTier 0 1.35 0.36
#> 8 s2T01.TextGrid 2 phones IntervalTier 0 1.35 0.495
#> 9 s2T01.TextGrid 2 phones IntervalTier 0 1.35 0.522
#> 10 s2T01.TextGrid 2 phones IntervalTier 0 1.35 0.621
#> xmax text annotation_num
#> <dbl> <chr> <int>
#> 1 0.297 "" 1
#> 2 0.522 "bird" 2
#> 3 0.972 "house" 3
#> 4 1.35 "" 4
#> 5 0.297 "sil" 1
#> 6 0.36 "B" 2
#> 7 0.495 "ER1" 3
#> 8 0.522 "D" 4
#> 9 0.621 "HH" 5
#> 10 0.783 "AW1" 6
#> # ℹ 140 more rowsBy default, read_textgrid() uses the file basename (the file-path minus the directory part) for the file column. But we can manually set the file value. Here, we use purrr::map2() to map the function over read_textgrid(path, file) over path and file pairs. Then we add the speaker information with some dataframe manipulation functions.
library(dplyr)
# This tells read_textgrid() to set the file column to the full path
data <- map2(paths, paths, read_textgrid) |>
list_rbind() |>
mutate(
# basename() removes the folder part from a path,
# dirname() removes the file part from a path
speaker = basename(dirname(file)),
file = basename(file),
) |>
select(
speaker, everything()
)
data
#> # A tibble: 150 × 11
#> speaker file tier_num tier_name tier_type tier_xmin tier_xmax
#> <chr> <chr> <int> <chr> <chr> <dbl> <dbl>
#> 1 speaker001 s2T01.TextGrid 1 words IntervalTier 0 1.35
#> 2 speaker001 s2T01.TextGrid 1 words IntervalTier 0 1.35
#> 3 speaker001 s2T01.TextGrid 1 words IntervalTier 0 1.35
#> 4 speaker001 s2T01.TextGrid 1 words IntervalTier 0 1.35
#> 5 speaker001 s2T01.TextGrid 2 phones IntervalTier 0 1.35
#> 6 speaker001 s2T01.TextGrid 2 phones IntervalTier 0 1.35
#> 7 speaker001 s2T01.TextGrid 2 phones IntervalTier 0 1.35
#> 8 speaker001 s2T01.TextGrid 2 phones IntervalTier 0 1.35
#> 9 speaker001 s2T01.TextGrid 2 phones IntervalTier 0 1.35
#> 10 speaker001 s2T01.TextGrid 2 phones IntervalTier 0 1.35
#> xmin xmax text annotation_num
#> <dbl> <dbl> <chr> <int>
#> 1 0 0.297 "" 1
#> 2 0.297 0.522 "bird" 2
#> 3 0.522 0.972 "house" 3
#> 4 0.972 1.35 "" 4
#> 5 0 0.297 "sil" 1
#> 6 0.297 0.36 "B" 2
#> 7 0.36 0.495 "ER1" 3
#> 8 0.495 0.522 "D" 4
#> 9 0.522 0.621 "HH" 5
#> 10 0.621 0.783 "AW1" 6
#> # ℹ 140 more rowsAnother strategy would be to read the textgrid dataframes into a list column and tidyr::unnest() them.
# Read dataframes into a list column
data_nested <- tibble(
speaker = basename(dirname(paths)),
data = map(paths, read_textgrid)
)
# We have one row per textgrid dataframe because `data` is a list column
data_nested
#> # A tibble: 10 × 2
#> speaker data
#> <chr> <list>
#> 1 speaker001 <tibble [13 × 10]>
#> 2 speaker001 <tibble [15 × 10]>
#> 3 speaker001 <tibble [16 × 10]>
#> 4 speaker001 <tibble [12 × 10]>
#> 5 speaker001 <tibble [19 × 10]>
#> 6 speaker002 <tibble [13 × 10]>
#> 7 speaker002 <tibble [15 × 10]>
#> 8 speaker002 <tibble [16 × 10]>
#> 9 speaker002 <tibble [12 × 10]>
#> 10 speaker002 <tibble [19 × 10]>
# promote the nested dataframes into the main dataframe
tidyr::unnest(data_nested, "data")
#> # A tibble: 150 × 11
#> speaker file tier_num tier_name tier_type tier_xmin tier_xmax xmin xmax
#> <chr> <chr> <int> <chr> <chr> <dbl> <dbl> <dbl> <dbl>
#> 1 speaker001 s2T0… 1 words Interval… 0 1.35 0 0.297
#> 2 speaker001 s2T0… 1 words Interval… 0 1.35 0.297 0.522
#> 3 speaker001 s2T0… 1 words Interval… 0 1.35 0.522 0.972
#> 4 speaker001 s2T0… 1 words Interval… 0 1.35 0.972 1.35
#> 5 speaker001 s2T0… 2 phones Interval… 0 1.35 0 0.297
#> 6 speaker001 s2T0… 2 phones Interval… 0 1.35 0.297 0.36
#> 7 speaker001 s2T0… 2 phones Interval… 0 1.35 0.36 0.495
#> 8 speaker001 s2T0… 2 phones Interval… 0 1.35 0.495 0.522
#> 9 speaker001 s2T0… 2 phones Interval… 0 1.35 0.522 0.621
#> 10 speaker001 s2T0… 2 phones Interval… 0 1.35 0.621 0.783
#> # ℹ 140 more rows
#> # ℹ 2 more variables: text <chr>, annotation_num <int>Pivoting nested intervals in textgrids
In the textgrids above, there is a natural nesting or hierarchy to the tiers. Intervals in words tier contain intervals in the phones tier. It is often necessary to group intervals by their parent intervals (group phones by words). This package provides the pivot_textgrid_tiers() function to convert textgrids into a wide format in a way that respects the nesting/hierarchy of tiers.
data_wide <- pivot_textgrid_tiers(
data,
tiers = c("words", "phones"),
join_cols = c("speaker", "file")
)
data_wide
#> # A tibble: 108 × 18
#> speaker file words words_xmin words_xmax words_xmid words_annotation_num
#> <chr> <chr> <chr> <dbl> <dbl> <dbl> <int>
#> 1 speaker001 s2T01… "" 0 0.297 0.149 1
#> 2 speaker001 s2T01… "bir… 0.297 0.522 0.410 2
#> 3 speaker001 s2T01… "bir… 0.297 0.522 0.410 2
#> 4 speaker001 s2T01… "bir… 0.297 0.522 0.410 2
#> 5 speaker001 s2T01… "hou… 0.522 0.972 0.747 3
#> 6 speaker001 s2T01… "hou… 0.522 0.972 0.747 3
#> 7 speaker001 s2T01… "hou… 0.522 0.972 0.747 3
#> 8 speaker001 s2T01… "" 0.972 1.35 1.16 4
#> 9 speaker001 s2T01… "" 0.972 1.35 1.16 4
#> 10 speaker001 s2T02… "" 0 0.297 0.149 1
#> # ℹ 98 more rows
#> # ℹ 11 more variables: words_tier_num <int>, words_tier_type <chr>,
#> # tier_xmin <dbl>, tier_xmax <dbl>, phones <chr>, phones_xmin <dbl>,
#> # phones_xmax <dbl>, phones_xmid <dbl>, phones_annotation_num <int>,
#> # phones_tier_num <int>, phones_tier_type <chr>
# more clearly
data_wide |>
select(
speaker, file, words, phones,
words_xmin, words_xmax, phones_xmin, phones_xmax
)
#> # A tibble: 108 × 8
#> speaker file words phones words_xmin words_xmax phones_xmin phones_xmax
#> <chr> <chr> <chr> <chr> <dbl> <dbl> <dbl> <dbl>
#> 1 speaker001 s2T01.… "" "sil" 0 0.297 0 0.297
#> 2 speaker001 s2T01.… "bir… "B" 0.297 0.522 0.297 0.36
#> 3 speaker001 s2T01.… "bir… "ER1" 0.297 0.522 0.36 0.495
#> 4 speaker001 s2T01.… "bir… "D" 0.297 0.522 0.495 0.522
#> 5 speaker001 s2T01.… "hou… "HH" 0.522 0.972 0.522 0.621
#> 6 speaker001 s2T01.… "hou… "AW1" 0.522 0.972 0.621 0.783
#> 7 speaker001 s2T01.… "hou… "S" 0.522 0.972 0.783 0.972
#> 8 speaker001 s2T01.… "" "sp" 0.972 1.35 0.972 1.33
#> 9 speaker001 s2T01.… "" "" 0.972 1.35 1.33 1.35
#> 10 speaker001 s2T02.… "" "sil" 0 0.297 0 0.297
#> # ℹ 98 more rowsSome remarks:
- Each tier in
tiersbecomes a batch of columns. For the rows for thewordstier become the batch of columnswords(the originaltextvalue),words_xmin,words_xmax, etc. - The columns in
join_colsshould uniquely identify a textgrid file, so the combination ofspeakerandfileis needed in the case where different speakers have the same file. - The tier names in
tiersshould be given in the order of their nesting from outside to inside (e.g.,wordscontainphones). Behind the scenes,dplyr::left_join(..., relationship = "one-to-many")is used to constrain how intervals are combined.
This function also works on a single tiers value. In this case, the function returns just the intervals in that tier with the columns renamed and prefixed.
data |>
pivot_textgrid_tiers(
tiers = "words",
join_cols = c("speaker", "file")
)
#> # A tibble: 42 × 11
#> speaker file words words_xmin words_xmax words_xmid words_annotation_num
#> <chr> <chr> <chr> <dbl> <dbl> <dbl> <int>
#> 1 speaker001 s2T01… "" 0 0.297 0.149 1
#> 2 speaker001 s2T01… "bir… 0.297 0.522 0.410 2
#> 3 speaker001 s2T01… "hou… 0.522 0.972 0.747 3
#> 4 speaker001 s2T01… "" 0.972 1.35 1.16 4
#> 5 speaker001 s2T02… "" 0 0.297 0.149 1
#> 6 speaker001 s2T02… "cow… 0.297 0.702 0.500 2
#> 7 speaker001 s2T02… "boo… 0.702 1.17 0.936 3
#> 8 speaker001 s2T02… "" 1.17 1.59 1.38 4
#> 9 speaker001 s2T03… "" 0 0.369 0.184 1
#> 10 speaker001 s2T03… "hug" 0.369 0.657 0.513 2
#> # ℹ 32 more rows
#> # ℹ 4 more variables: words_tier_num <int>, words_tier_type <chr>,
#> # tier_xmin <dbl>, tier_xmax <dbl>Speeding things up
Do you have thousands of textgrids to read? The following workflow can speed things up. We are going to read the textgrids in parallel. Below are two approaches:
- future backend and furrr frontend
- mirai backend and purrr frontend
The backend manages the parallel computation, and the frontend provides the syntax for calling a function with parallelism.
Approach 1: We tell future to use a multisession plan for parallelism, so the computations are done on separate R sessions in the background. The syntax is like the above purrr code, but we replace map() with future_map().
library(future)
library(furrr)
plan(multisession, workers = 4)
data_nested <- tibble(
speaker = basename(dirname(paths)),
data = future_map(paths, read_textgrid)
)Approach 2: We have mirai set up 4 daemons (background processes), and then we use purrr’s in_parallel() helper to signal to map() that the function should be run in parallel. We need to give all the information needed for the daemons to run the function, so we 1) provide a complete function definition (including function(x) ...) and 2) spell out the package namespace readtextgrid::read_textgrid().
mirai::daemons(4)
data_nested <- tibble(
speaker = basename(dirname(paths)),
data = map(paths, in_parallel(function(x) readtextgrid::read_textgrid(x)))
)
mirai::daemons(0)Another way to eke out performance is to set the encoding. By default, readtextgrid uses readr::guess_encoding() to determine the encoding of the textgrid before reading it in. But if you know the encoding beforehand, you can skip this guessing. In my limited testing, I found that setting the encoding could reduce benchmark times by 3–4% compared to guessing the encoding.
Here, we read 100 textgrids using different approaches to benchmark the results.
paths_bench <- withr::with_seed(1, sample(paths, 100, replace = TRUE))
mirai::daemons(4)
bench::mark(
lapply_guess = lapply(paths_bench, read_textgrid),
lapply_set = lapply(paths_bench, read_textgrid, encoding = "UTF-8"),
future_guess = future_map(paths_bench, read_textgrid),
future_set = future_map(paths_bench, read_textgrid, encoding = "UTF-8"),
mirai_guess = purrr::map(
paths_bench,
in_parallel(function(x) readtextgrid::read_textgrid(x))
),
mirai_set = purrr::map(
paths_bench,
in_parallel(function(x) readtextgrid::read_textgrid(x, encoding = "UTF-8"))
),
check = TRUE
)
#> Warning: Some expressions had a GC in every iteration; so filtering is
#> disabled.
#> # A tibble: 6 × 6
#> expression min median `itr/sec` mem_alloc `gc/sec`
#> <bch:expr> <bch:tm> <bch:tm> <dbl> <bch:byt> <dbl>
#> 1 lapply_guess 1.17s 1.17s 0.852 13.32MB 5.96
#> 2 lapply_set 883.69ms 883.69ms 1.13 5.41MB 6.79
#> 3 future_guess 407.83ms 421.37ms 2.37 627.53KB 2.37
#> 4 future_set 356.49ms 358.09ms 2.79 627.53KB 2.79
#> 5 mirai_guess 315.85ms 338.82ms 2.95 1006.66KB 0
#> 6 mirai_set 258.63ms 259.45ms 3.85 1006.66KB 0
mirai::daemons(0)Legacy behavior and supported textgrid formats
The original version of this package assumed the textgrid text files followed a “long” format with helpful labels and annotations. For example, in the following textgrid, each number has a label that makes it easy and fast to parse the textgrid with regular expressions:
File type = "ooTextFile"
Object class = "TextGrid"
xmin = 0
xmax = 1
tiers? <exists>
size = 1
item []:
item [1]:
class = "IntervalTier"
name = "Mary"
xmin = 0
xmax = 1
intervals: size = 1
intervals [1]:
xmin = 0
xmax = 1
text = "" The original version of the parser designed for this textgrid format is still provided with the legacy_read_textgrid() and legacy_read_textgrid_lines() functions.
Version 0.2.0 of readtextgrid added a C++ based parser that can handle many more textgrid formats. For example, it can “short” format textgrids like the following:
File type = "ooTextFile"
Object class = "TextGrid"
0
1
<exists>
1
"IntervalTier"
"Mary"
0
1
1
0
1
""The “long” format textgrids are outputted in Praat with Save > Save as text file..., and the “short” format textgrids are outputted with Save > Save as short textfile....
readtextgrid’s parser can also handle esoteric features like comments (that start with !) or arbitrary text attached to a number, as in the following example;:
File type = "ooTextFile"
Object class = "TextGrid"
! info about the grid
0s 1s <exists> 1
! info about the tier
"IntervalTier" "Mary" 0s 1s 1 ! type, name, xmin, xmax, size
0s 1s "" ! interval xmin, xmax, sizeBecause the new parser uses C++ for tokenization—that is, the part scans the contents character by character and determines whether the inputs are strings, numbers, or skipped—it is much faster the legacy version.
paths_bench <- withr::with_seed(2, sample(paths, 10, replace = TRUE))
bench::mark(
current = lapply(paths_bench, read_textgrid),
legacy = lapply(paths_bench, legacy_read_textgrid),
min_iterations = 10,
filter_gc = FALSE,
check = TRUE
)
#> # A tibble: 2 × 6
#> expression min median `itr/sec` mem_alloc `gc/sec`
#> <bch:expr> <bch:tm> <bch:tm> <dbl> <bch:byt> <dbl>
#> 1 current 114ms 123ms 8.11 1.31MB 4.86
#> 2 legacy 332ms 342ms 2.89 19.57MB 6.06Other tips
Helpful columns
The following columns are often helpful:
-
durationof an interval -
xmidmidpoint of an interval -
total_annotationstotal number of annotations on a tier
Here is how to create them:
data |>
# grouping needed for counting annotations per tier per file per speaker
group_by(speaker, file, tier_num) |>
mutate(
duration = xmax - xmin,
xmid = xmin + (xmax - xmin) / 2,
total_annotations = sum(!is.na(annotation_num))
) |>
ungroup() |>
glimpse()
#> Rows: 150
#> Columns: 14
#> $ speaker <chr> "speaker001", "speaker001", "speaker001", "speaker00…
#> $ file <chr> "s2T01.TextGrid", "s2T01.TextGrid", "s2T01.TextGrid"…
#> $ tier_num <int> 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 2…
#> $ tier_name <chr> "words", "words", "words", "words", "phones", "phone…
#> $ tier_type <chr> "IntervalTier", "IntervalTier", "IntervalTier", "Int…
#> $ tier_xmin <dbl> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
#> $ tier_xmax <dbl> 1.348571, 1.348571, 1.348571, 1.348571, 1.348571, 1.…
#> $ xmin <dbl> 0.000, 0.297, 0.522, 0.972, 0.000, 0.297, 0.360, 0.4…
#> $ xmax <dbl> 0.297000, 0.522000, 0.972000, 1.348571, 0.297000, 0.…
#> $ text <chr> "", "bird", "house", "", "sil", "B", "ER1", "D", "HH…
#> $ annotation_num <int> 1, 2, 3, 4, 1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 2, 3, 4, 1…
#> $ duration <dbl> 0.29700000, 0.22500000, 0.45000000, 0.37657143, 0.29…
#> $ xmid <dbl> 0.148500, 0.409500, 0.747000, 1.160286, 0.148500, 0.…
#> $ total_annotations <int> 4, 4, 4, 4, 9, 9, 9, 9, 9, 9, 9, 9, 9, 4, 4, 4, 4, 1…Launching Praat
This tip is written from the perspective of a Windows user who uses git bash for a terminal.
To open textgrids in Praat, you can tell R to call Praat from the command line. You have to know where the location of the Praat binary is though. I like to keep a copy in my project directories. So, assuming that Praat.exe in my working folder, the following would open the 10 textgrids in paths in Praat.
Acknowledgments
readtextgrid was created to process data from the WISC Lab project. Thus, development of this package was supported by NIH R01DC009411 and NIH R01DC015653.
Please note that the ‘readtextgrid’ project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.