aeneas Library Tutorial

Overview

Although a majority of aeneas users work with the built-in command line tools, aeneas is primarily designed for being used as a Python library. Even the aeneas.tools can be used programmatically, thanks to their standard I/O interface.

Example

Create a Task and process it, outputting the resulting sync map to file:

#!/usr/bin/env python
# coding=utf-8

from aeneas.executetask import ExecuteTask
from aeneas.task import Task

# create Task object
config_string = u"task_language=eng|is_text_type=plain|os_task_file_format=json"
task = Task(config_string=config_string)
task.audio_file_path_absolute = u"/path/to/input/audio.mp3"
task.text_file_path_absolute = u"/path/to/input/plain.txt"
task.sync_map_file_path_absolute = u"/path/to/output/syncmap.json"

# process Task
ExecuteTask(task).execute()

# output sync map to file
task.output_sync_map_file()

You can also use ExecuteTaskCLI:

#!/usr/bin/env python
# coding=utf-8

from aeneas.tools.execute_task import ExecuteTaskCLI

ExecuteTaskCLI(use_sys=False).run(arguments=[
    None, # dummy program name argument
    u"/path/to/input/audio.mp3",
    u"/path/to/input/plain.txt",
    u"task_language=eng|is_text_type=plain|os_task_file_format=json",
    u"/path/to/output/syncmap.json"
])

Clearly, you can also manipulate objects programmatically.

Example

Create a Task, process it, and print all fragments in the resulting sync map whose duration is less than five seconds:

#!/usr/bin/env python
# coding=utf-8

from aeneas.executetask import ExecuteTask
from aeneas.task import Task

# create Task object
config_string = u"task_language=eng|is_text_type=plain|os_task_file_format=json"
task = Task(config_string=config_string)
task.audio_file_path_absolute = u"/path/to/input/audio.mp3"
task.text_file_path_absolute = u"/path/to/input/plain.txt"

# process Task
ExecuteTask(task).execute()

# print fragments with a duration < 5 seconds
for fragment in task.sync_map_leaves():
    if fragment.length < 5.0:
        print(fragment)

Instead of passing around configuration strings, you can set properties explicitly, using the library functions and constants.

Example

Create a Task, process it, and print the resulting sync map:

#!/usr/bin/env python
# coding=utf-8

from aeneas.exacttiming import TimeValue
from aeneas.executetask import ExecuteTask
from aeneas.language import Language
from aeneas.syncmap import SyncMapFormat
from aeneas.task import Task
from aeneas.task import TaskConfiguration
from aeneas.textfile import TextFileFormat
import aeneas.globalconstants as gc

# create Task object
config = TaskConfiguration()
config[gc.PPN_TASK_LANGUAGE] = Language.ENG
config[gc.PPN_TASK_IS_TEXT_FILE_FORMAT] = TextFileFormat.PLAIN
config[gc.PPN_TASK_OS_FILE_FORMAT] = SyncMapFormat.JSON
task = Task()
task.configuration = config
task.audio_file_path_absolute = u"/path/to/input/audio.mp3"
task.text_file_path_absolute = u"/path/to/input/plain.txt"

# process Task
ExecuteTask(task).execute()

# print produced sync map
print(task.sync_map)

Dependencies

  • numpy (v1.9 or later)
  • lxml (v3.6.0 or later)
  • BeautifulSoup (v4.5.1 or later)

Only numpy is actually needed, as it is heavily used for the alignment computation.

The other two dependencies (lxml and BeautifulSoup) are needed only if you use XML-like input or output formats. However, since they are popular Python packages, to avoid complex import testing, they are listed as requirements. This choice might change in the future.

Depending on what aeneas classes you want to use, you might need to install the following optional dependencies:

  • boto3 (for using the AWS Polly TTS API wrapper)
  • requests (for using the Nuance TTS API wrapper)
  • Pillow (for plotting waveforms with plotter)
  • tgt (for outputting sync maps to TextGrid format)
  • youtube-dl and pafy (for downloading audio from Internet with Downloader)

Speeding Critical Sections Up: Python C/C++ Extensions

Forced alignment is a computationally demanding task, both CPU-intensive and memory-intensive. Aligning a dozen minutes of audio might require an hour if done with pure Python code.

Hence, critical sections of the alignment code are written as Python C/C++ extensions, that is, C/C++ code that receives input from Python code, performs the heavy computation, and returns results to the Python code. The rule of thumb is that the C/C++ code only perform “computation-like”, low-level functions, while “house-keeping”, high-level functions are done in Python land.

With this approach, aligning a dozen minutes of audio requires only few seconds, and even aligning hours of audio can be done in few minutes. The drawback is that your environment must be able to compile Python C/C++ extensions. If you install aeneas via PyPI (e.g., pip install aeneas), the compilation step is done automatically for you.

Warning

Due to the Python C/C++ extension compile and setup mechanism, you must install numpy before installing aeneas, and there is no (sane) way for the aeneas setup.py to install numpy before compiling the aeneas source code. Hence, you really need to (manually) install numpy before installing aeneas. Hopefully this inconvenience will be removed in the future.

The Python C/C++ extensions included in aeneas are:

  • aeneas.cdtw, for computing the DTW;
  • aeneas.cew, for synthesizing text via the eSpeak C API;
  • aeneas.cfw, for synthesizing text via the Festival C++ API;
  • aeneas.cmfcc, for computing a MFCC representation of a WAVE (RIFF) audio file;
  • aeneas.cwave, for reading WAVE (RIFF) audio files.

Note

Currently aeneas.cew is available on Linux, Mac OS X, and Windows. On Windows 64 bit it does not seem to work, probably because eSpeak is available only as a 32 bit program/library, and hence aeneas will fall back to run the pure Python code. Starting with v1.5.0, the pure Python code for synthesizing text with eSpeak via subprocess is only 2-3 times slower than aeneas.cew. Unless you work with thousands of text fragments, the performance difference is negligible.

Note

Currently aeneas.cfw is experimental and disabled by default. Probably it works only on Linux. To compile it, make sure you have installed the Festival and speech_tools libraries (e.g., install the festival-dev package on DEB-based OSes) and set the environment variable AENEAS_FORCE_CFW=True before running pip install aeneas or python setup.py.

Note

Currently aeneas.cwave is not used. It will be enabled in a future version of aeneas.

Concepts

Except for “enumeration” classes (e.g., TextFileFormat) and “data-only” classes (e.g., TextFragment), most classes are subclasses of Loggable, which provides the ability to log events using a shared Logger object (logger), and to inject runtime execution parameters using a shared RuntimeConfiguration object (rconf).

The logger can tee (i.e., store messages and print them to stdout) or dump to file.

The rconf provides a way to fine tune aeneas by changing its internal behavior. The library defaults should fine for most use cases, and they do not require explicitly passing an rconf object.

Example

Process a task with custom parameters, and log messages:

# create Logger which logs and tees
logger = Logger(tee=True)

# create RuntimeConfiguration object, with custom MFCC length and shift
rconf = RuntimeConfiguration()
rconf[RuntimeConfiguration.MFCC_WINDOW_LENGTH] = TimeValue(u"0.150")
rconf[RuntimeConfiguration.MFCC_WINDOW_SHIFT] = TimeValue(u"0.050")

# create Task object
task = ...

# process Task with custom parameters
ExecuteTask(task, rconf=rconf, logger=logger).execute()

If you read from/write to file, you should be fine interacting only with Task functions. For example, setting a path in audio_file_path_absolute() (resp., text_file_path_absolute()) force the library to load the given file, and to create a AudioFile (resp., TextFile) object behind the scenes, storing it inside the Task object.

However, you can also build e.g. your own TextFile and then assign it to your Task.

Example

Create a TextFile programmatically, and assign it to Task:

task = Task()
textfile = TextFile()
for identifier, frag_text in [
    (u"f001", [u"first fragment"]),
    (u"f002", [u"second fragment"]),
    (u"f003", [u"third fragment"])
]:
    textfile.add_fragment(TextFragment(identifier, Language.ENG, frag_text, frag_text))
task.text_file = textfile

Starting with v1.5.0, both TextFile and SyncMap are backed by the Tree structure, which can represent multilevel I/O files. Both have a “virtual” (empty) root node, to which the “level 1” nodes are attached. Note that single-level text files and sync maps are a special case, where only “level 1” nodes are present, producing a tree with a root node and a list of children, effectively equivalent to the “list” structure pre-v1.5.0.

Miscellanea

  • Ensuring that all the strings you pass to aeneas are Unicode strings will save you a lot of headaches. If you read from files, be sure they are encoded using UTF-8.
  • You can use any audio file format that is supported by ffprobe and ffmpeg. If unsure, just try to play them on your audio file on the console: if it works there, it should work inside aeneas too.
  • Enumeration classes usually have an ALLOWED_VALUE class member, which lists all the allowed values. For example: ALLOWED_VALUES. This list is used for example by the validator to check input values.
  • Most classes are optimized for reducing memory consumption. For example, if you create an AudioFileMFCC with a file path, the input audio file will be converted to a temporary WAVE file, audio samples will be read into memory, MFCCs will be computed, and then audio data will be discarded from memory and the temporary WAVE file will be deleted, keeping only the MFCC matrix into memory. If you prefer persistence, you need to build intermediate objects yourself (i.e., FFMPEGWrapper, AudioFile, etc.) and properly dispose of them in your code.
  • Wherever possible, NumPy views are used to avoid data copying. Similarly, built-in NumPy functions are used to improve run time.
  • To avoid numerical issues, always use TimeValue to hold time values with arbitrary precision. Note that doing so incurs in a negligible execution slow down, because the heaviest computations are done with integer NumPy indices and arrays and the transformation to TimeValue takes place only when the sync map is output to file.

Package aeneas

The main aeneas package contains several subpackages:

and the following modules:

Package aeneas.extra

The aeneas.extra package contains some extra Python source files which provide experimental and not officially supported functions, mainly custom, not built-in TTS engine wrappers.

For example, if you want to write your own custom TTS engine wrapper, have a look at the aeneas/extra/ctw_espeak.py source file, which is heavily commented and should be easy to modify for your own TTS engine.

Package aeneas.tests

The aeneas.tests package contains the unit test files for aeneas.

Resources needed to run the tests, for example audio and text files, are located in the aeneas/tests/res/ directory.

Package aeneas.tools

The aeneas.tools package contains the built-in command line tools for aeneas.

The two main tools are:

  • aeneas.tools.execute_job
  • aeneas.tools.execute_task

which are described in the aeneas Built-in Command Line Tools Tutorial.

Moreover, the aeneas.tools package also contains the following programs, useful for debugging or converting between different file formats:

  • aeneas.tools.convert_syncmap: convert a sync map from a format to another
  • aeneas.tools.download: download a file from a Web resource (currently, audio from a YouTube video)
  • aeneas.tools.extract_mfcc: extract MFCCs from a monoaural WAVE file
  • aeneas.tools.ffmpeg_wrapper: a wrapper around ffmpeg
  • aeneas.tools.ffprobe_wrapper: a wrapper around ffprobe
  • aeneas.tools.plot_waveform: plot a waveform and sets of labels to file
  • aeneas.tools.read_audio: read the properties of an audio file
  • aeneas.tools.read_text: read a text file and show the extracted text fragments
  • aeneas.tools.run_sd: read an audio file and the corresponding text file and detect the audio head/tail
  • aeneas.tools.run_vad: read an audio file and compute speech/nonspeech time intervals
  • aeneas.tools.synthesize_text: synthesize several text fragments read from file into a single wav file
  • aeneas.tools.validate: validate a job container or configuration strings/files

Run each program without arguments to get its help manual and usage examples.

Resources needed to run the live examples, for example audio and text files, are located in the aeneas/tools/res/ directory.

The package also contains the aeneas.tools.hydra script, which can run any of the tools listed above. Run it without arguments to get its manual.

Package aeneas.ttswrappers

The aeneas.ttswrappers package contains the wrappers for several built-in TTS engines which can be used in the synthesis step of the alignment procedure.