A pure python module for reading and writing kaldi ark files
kaldiio is an IO utility implemented in pure Python language for several file formats used in kaldi, which are named asark and scp. ark and scp are used in in order to archive some objects defined in Kaldi, typically it is Matrix object of Kaldi.
In this section, we describe the basic concept of ark and scp. More detail about the File-IO in Kaldi-asr: http://kaldi-asr.org/doc/io.html
ark is an archive format to save any Kaldi objects. This library mainly support KaldiMatrix/KaldiVector.
This ia an example of ark file of KaldiMatrix: ark file
If you have Kaldi, you can convert it to text format as following
# copy-feats <read-specifier> <write-specifier>
copy-feats ark:test.ark ark,t:text.arkcopy-feats is designed to have high affinity with unix command line:
-
arkcan be flushed to and from unix pipe.cat test.ark | copy-feats ark:- ark,t:- | less # Show the contents in the ark-indicates standard input stream or output stream. -
Unix command can be used as
read-specifierandwspecifiercopy-feats ark:'gunzip -c some.ark.gz |' ark:some.ark
scp is a text file such as,
uttid1 /some/where/feats.ark:123
uttid2 /some/where/feats.ark:156
uttid3 /some/where/feats.ark:245
The first column, uttid1, indicates the utterance id and the second, /some/where/feats.ark:123, is the file path of matrix/vector of kaldi formats. The number after colon is a starting addressof the object of the file.
scp looks very simple format, but has several powerful features.
-
Mutual conversion between
arkandscpcopy-feats scp:foo.scp ark:foo.ark # scp -> ark copy-feats ark:foo.ark ark,scp:bar.ark,bar.scp # ark -> ark,scp -
Unix command can be used insead of direct file path
For example, the following scp file can be also used.
uttid1 cat /some/where/feats1.mat | uttid2 cat /some/where/feats2.mat | uttid3 cat /some/where/feats3.mat |
wav.scp is a scp to describe wave file paths.
uttid1 /some/path/a.wav
uttid2 /some/path/b.wav
uttid3 /some/path/c.wav
wav.scp is also can be embeded unix command as normal scp file. This is often used for converting file format in kaldi recipes.
uttid1 sph2pipe -f wav /some/path/a.wv1 |
uttid2 sph2pipe -f wav /some/path/b.wv1 |
uttid3 sph2pipe -f wav /some/path/c.wv1 |
Kaldiio supports:
- Read/Write for archive formats: ark, scp
- Binary/Text - Float/Double Matrix: DM, FM
- Binary/Text - Float/Double Vector: DV, FV
- Compressed Matrix for loading: CM, CM2, CM3
- Compressed Matrix for writing: All compressoin_method are supported: 1,2,3,4,5,6,7
- Binary/Text for Int-vector, typically used for
alifiles.
- Read/Write via a pipe: e.g. "ark: cat feats.ark |"
- Read wav.scp / wav.ark
- (New!) Some extended ark format not supported in Kaldi originally.
- The ark file for numpy, pickle, wav, flac files.
The followings are not supported
- Write in existing scp file
- NNet2/NNet3 egs
- Lattice file
- Python-C++ binding
- https://github.com/pykaldi/pykaldi
- Looks great. I recommend pykaldi if you aren't particular about pure python.
- https://github.com/janchorowski/kaldi-python/
- Maybe not enough maintained now.
- https://github.com/t13m/kaldi-readers-for-tensorflow
- Ark reader for tensorflow
- https://github.com/csukuangfj/kaldi_native_io
- Implemented in C++
- Have interface for Python
- Support all types of
rspecifierandwspecifier - Have a uniform interface for writing, sequential reading, and random access reading
pip install kaldi_native_io
- https://github.com/pykaldi/pykaldi
- Pure Python
- https://github.com/vesis84/kaldi-io-for-python
kaldiiois based on this module, butkaldiiosupports more features than it.
- https://github.com/funcwj/kaldi-python-io
- Python>=3.6.
nnet3-egsis also supported.
- Python>=3.6.
- https://github.com/vesis84/kaldi-io-for-python
pip install kaldiiokaldiio doesn't distinguish the API for each kaldi-objects, i.e.
Kaldi-Matrix, Kaldi-Vector, not depending on whether it is binary or text, or compressed or not,
can be handled by the same API.
ReadHelper supports sequential accessing for scp or ark. If you need to access randomly, then use kaldiio.load_scp.
- Read matrix-scp
from kaldiio import ReadHelper
with ReadHelper('scp:file.scp') as reader:
for key, numpy_array in reader:
...- Read gziped ark
from kaldiio import ReadHelper
with ReadHelper('ark: gunzip -c file.ark.gz |') as reader:
for key, numpy_array in reader:
...
# Ali file
with ReadHelper('ark: gunzip -c exp/tri3_ali/ali.*.gz |') as reader:
for key, numpy_array in reader:
...- Read wav.scp
from kaldiio import ReadHelper
with ReadHelper('scp:wav.scp') as reader:
for key, (rate, numpy_array) in reader:
...γγγγ- v2.11.0: Removed wav option. You can load wav.scp without any addtional argument.
- Read wav.scp with segments
from kaldiio import ReadHelper
with ReadHelper('scp:wav.scp', segments='segments') as reader
for key, (rate, numpy_array) in reader:
...- Read from stdin
from kaldiio import ReadHelper
with ReadHelper('ark:-') as reader:
for key, numpy_array in reader:
...- Write matrices and vectors in a ark with scp
import numpy
from kaldiio import WriteHelper
with WriteHelper('ark,scp:file.ark,file.scp') as writer:
for i in range(10):
writer(str(i), numpy.random.randn(10, 10))
# The following is equivalent
# writer[str(i)] = numpy.random.randn(10, 10)- Write in compressed matrix
import numpy
from kaldiio import WriteHelper
with WriteHelper('ark:file.ark', compression_method=2) as writer:
for i in range(10):
writer(str(i), numpy.random.randn(10, 10))- Write matrices in text
import numpy
from kaldiio import WriteHelper
with WriteHelper('ark,t:file.ark') as writer:
for i in range(10):
writer(str(i), numpy.random.randn(10, 10))- Write in gziped ark
import numpy
from kaldiio import WriteHelper
with WriteHelper('ark:| gzip -c > file.ark.gz') as writer:
for i in range(10):
writer(str(i), numpy.random.randn(10, 10))- Write matrice to stdout
import numpy
from kaldiio import WriteHelper
with WriteHelper('ark:-') as writer:
for i in range(10):
writer(str(i), numpy.random.randn(10, 10))- (New!) Extended ark format using numpy, pickle, soundfile
import numpy
from kaldiio import WriteHelper
# NPY ARK
with WriteHelper('ark:-', write_function="numpy") as writer:
writer("foo", numpy.random.randn(10, 10))
# PICKLE ARK
with WriteHelper('ark:-', write_function="pickle") as writer:
writer("foo", numpy.random.randn(10, 10))
# FLAC ARK
with WriteHelper('ark:-', write_function="soundfile_flac") as writer:
writer("foo", numpy.random.randn(1000))Note that soundfile is an optional module and you need to install it to use this feature.
pip install soundfileWriteHelper and ReadHelper are high level wrapper of the following API to support kaldi style arguments.
import kaldiio
d = kaldiio.load_ark('a.ark') # d is a generator object
for key, numpy_array in d:
...
# === load_ark can accepts file descriptor, too
with open('a.ark') as fd:
for key, numpy_array in kaldiio.load_ark(fd):
...
# === Use with open_like_kaldi
from kaldiio import open_like_kaldi
with open_like_kaldi('gunzip -c file.ark.gz |', 'r') as f:
for key, numpy_array in kaldiio.load_ark(fd):
...load_arkcan load both matrices of ark and vectors of ark and also, it can be both text and binary.
load_scp creates "lazy dict", i.e.
The data are loaded in memory when accessing the element.
import kaldiio
d = kaldiio.load_scp('a.scp')
for key in d:
numpy_array = d[key]
with open('a.scp') as fd:
kaldiio.load_scp(fd)
d = kaldiio.load_scp('data/train/wav.scp', segments='data/train/segments')
for key in d:
rate, numpy_array = d[key]The object created by load_scp is a dict-like object, thus it has methods of dict.
import kaldiio
d = kaldiio.load_scp('a.scp')
d.keys()
d.items()
d.values()
'uttid' in d
d.get('uttid')load_scp_sequential creates "generator" as same as load_ark.
If you don't need random-accessing for each elements
and use it just to iterate for whole data,
then this method possibly performs faster than load_scp.
import kaldiio
d = kaldiio.load_scp_sequential('a.scp')
for key, numpy_array in d:
...d = kaldiio.load_scp('wav.scp')
for key in d:
rate, numpy_array = d[key]
# Supporting "segments"
d = kaldiio.load_scp('data/train/wav.scp', segments='data/train/segments')
for key in d:
rate, numpy_array = d[key]- v2.11.0:
load_wav_scpis deprecated now. Useload_scp.
numpy_array = kaldiio.load_mat('a.mat')
numpy_array = kaldiio.load_mat('a.ark:1134') # Seek and load
# If the file is wav, gets Tuple[int, numpy.ndarray]
rate, numpy_array = kaldiio.load_mat('a.wav')load_matcan load kaldi-matrix, kaldi-vector, and wave
# === Create ark file from numpy
kaldiio.save_ark('b.ark', {'key': numpy_array, 'key2': numpy_array2})
# Create ark with scp _file, too
kaldiio.save_ark('b.ark', {'key': numpy_array, 'key2': numpy_array2},
scp='b.scp')
# === Writes arrays to sys.stdout
import sys
kaldiio.save_ark(sys.stdout, {'key': numpy_array})
# === Writes arrays for each keys
# generate a.ark
kaldiio.save_ark('a.ark', {'key': numpy_array, 'key2': numpy_array2})
# After here, a.ark is opened with 'a' (append) mode.
kaldiio.save_ark('a.ark', {'key3': numpy_array3}, append=True)
# === Use with open_like_kaldi
from kaldiio import open_like_kaldi
with open_like_kaldi('| gzip a.ark.gz', 'w') as f:
kaldiio.save_ark(f, {'key': numpy_array})
kaldiio.save_ark(f, {'key2': numpy_array2})# array.ndim must be 1 or 2
kaldiio.save_mat('a.mat', numpy_array)save_matcan save both kaldi-matrix and kaldi-vector
kaldiio.open_like_kaldi is a useful tool if you are familiar with Kaldi. This function can performs as following,
from kaldiio import open_like_kaldi
with open_like_kaldi('echo -n hello |', 'r') as f:
assert f.read() == 'hello'
with open_like_kaldi('| cat > out.txt', 'w') as f:
f.write('hello')
with open('out.txt', 'r') as f:
assert f.read() == 'hello'
import sys
with open_like_kaldi('-', 'r') as f:
assert f is sys.stdin
with open_like_kaldi('-', 'w') as f:
assert f is sys.stdoutFor example, if there are gziped alignment file, then you can load it as:
from kaldiio import open_like_kaldi, load_ark
with open_like_kaldi('gunzip -c exp/tri3_ali/ali.*.gz |', 'rb') as f:
# Alignment format equals ark of IntVector
g = load_ark(f)
for k, numpy_array in g:
...from kaldiio import parse_specifier, open_like_kaldi, load_ark
rspecifier = 'ark:gunzip -c file.ark.gz |'
spec_dict = parse_specifier(rspecifier)
# spec_dict = {'ark': 'gunzip -c file.ark.gz |'}
with open_like_kaldi(spec_dict['ark'], 'rb') as fark:
for key, numpy_array in load_ark(fark):
...