search

TalkBank / batchalign2

Tools for language sample analysis.
https://talkbank.org/info/BA2-usage.pdf
BSD 3-Clause "New" or "Revised" License
15 stars 4 forks source link

readme

TalkBank | Batchalign2

Welcome! Batchalign2 is a Python suite of language sample analysis (LSA) software from the TalkBank project. It is used to interact with conversation audio files and their transcripts, and provides a whole host of analyses within this space.

The TalkBank Project, of which Batchalign is a part, is supported by NIH grant HD082736.

Quick Start

The following instructions provide a quick start to installing Batchalign. For most users aiming to process CHAT and audio with Batchalign, we recommend more detailed usage instructions: for usage and human transcript cleanup. The following provides a quick start guide for the program.

Get Python

Install and Update the Package

You can get Batchalign from PyPi, and you can update the package in the same way:

macOS/Linux:

pip install -U batchalign

Windows:

py -m pip install -U batchalign

Rock and Roll

There are two main ways of interacting with Batchalign. Batchalign can be used as a program to batch-process CHAT (hence the name), or as a Python LSA library.

Quick Start: Command Line

Basic Usage

Once installed, you can invoke the Batchalign program by typing batchalign into the Terminal (MacOS) or Command Prompt (Windows).

It is used in the following basic way:

batchalign [verb] [input_dir] [output_dir]

Where verb includes:

  1. transcribe - by placing only an audio of video file (.mp3/.mp4/.wav) in the input directory, this function performs ASR on the audio, diarizes utterances, identifies some basic conversational features like retracing and filled pauses, and generates word-level alignments. You must supply a language code flag: --lang=[three letter ISO language code] for the ASR system to know what language the transcript is in. You can choose the flags --rev to use Rev.AI, a commercial ASR service, or --whisper, to use a local copy of OpenAI Whisper.
  2. align - by placing both an audio of video file (.mp3/.mp4/.wav) and an utterance-aligned CHAT file in the input directory, this function recovers utterance-level time alignments (if they are not already annotated) and generates word-level alignments. The @Languages header in the CHAT file tells the program which language is in the transcript.
  3. morphotag - by placing a CHAT file in the input directory, this function uses Stanford NLP Stanza to generate morphological and dependency analyses. The @Languages header in the CHAT file tells the program which language is in the transcript. You must supply a language code flag: --lang=[three letter ISO language code] for the alignment system to know what language the transcript is in.

You can get a CHAT transcript to experiment with at the TalkBank website, under any of the "Banks" that are available. You can also generate and parse a CHAT transcript via the Python program.

Sample Commands

For input files (CHAT and audio for align, CHAT only for morphotag, and audio only for transcribe), located in ~/ba_input dumping the output to ~/ba_output, one could write:

ASR + Segmentation

batchalign transcribe --lang=eng ~/ba_input ~/ba_output

morphosyntactic analysis

batchalign morphotag ~/ba_input ~/ba_output

forced alignment

batchalign align ~/ba_input ~/ba_output

Follow instructions from

batchalign --help

and 

batchalign [verb] --help

to learn more about other options.

Verbosity

Placing one or multiple -v behind the word batchalign (i.e. behind the [verb] will not work) increases the verbosity of Batchalign. The default mode and one -v will use the normal Batchalign interface, whereas Batchalign with more than 1 -v will switch to the text-based "logging" interface.

For instance, here is the instruction for running Batchalign to perform forced-alignment:

batchalign align input output

With one -v, you can get stack trace information about any files that crashes: 

batchalign -v align input output

and with two -vv, we will ditch the loading bar user interface and instead switch to a logging-based interface that has more information about what Batchalign is doing under the hood:

batchalign -vv align input output

Quick Start: Python

Let's begin!

import batchalign as ba

Document

The Document is the most basic object in Bachalign. All processing pipelines expect Document as input, and will spit out Document as output.

doc = ba.Document.new("Hello, this is a transcript! I have two utterances.", 
                      media_path="audio.mp3", lang="eng")

# navigating the document
first_utterance = doc[0]
first_form = doc[0][0]
the_comma = doc[0][1]

assert the_comma.text == ','
assert the_comma.type == ba.TokenType.PUNCT

# taking a transcript
sentences = doc.transcript(include_tiers=False, strip=True)

Notably, if you have a Document that you haven't transcribed yet, you still can make a Document!

doc = ba.Document.new(media_path="audio.mp3", lang="eng")

Pipelines

Quick Pipeline

Say you wanted to perform ASR, and then tag morphology of the resulting output.

nlp = ba.BatchalignPipeline.new("asr,morphosyntax", lang="eng", num_speakers=2)
doc = ba.Document.new(media_path="audio.mp3", lang="eng")
doc = nlp(doc) # this is equivalent to nlp("audio.mp3"), we will make the initial doc for you

first_word_pos = doc[0][0].morphology
first_word_time = doc[0][0].time
first_utterance_time = doc[0].alignment

The quick API (right now) has support for the following tasks, which you can pass in a comma-separated list in the first argument:

We will support many, many, many more tasks soon with this API. For now, to gain access to the whole suite of tools, use the second pipeline API discussed below.

Manual Pipeline

Batchalign ships with a plurality of engines which preform the actual processing. For instance, to recreate the demo we had above using the Engines API, we would write

# ASR
whisper = ba.WhisperEngine(lang="eng")
# retracing and disfluency analysis
retrace = ba.NgramRetraceEngine()
disfluency = ba.DisfluencyReplacementEngine()
# morphosyntax
morphosyntax = ba.StanzaEngine()

# create a pipeline
nlp = ba.BatchalignPipeline(whisper, retrace, disfluency, morphosyntax)

# and run it!                             
doc = nlp("audio.mp3")

Here's a list of available engines.

Formats

We currently support reading and writing two transcript formats: TalkBank CHAT, and Praat TextGrid.

CHAT

Here's how to read and write a CHAT file to parse a TalkBank transcript!

# reading
chat = ba.CHATFile(path="chat.cha")
doc = chat.doc

# writing
chat = ba.CHATFile(doc=doc)
chat.write("chat.cha")

We will automatically detect audio files located within the same directory as the CHAT file, and associate it with the Batchalign Document.

TextGrid

Importantly, there are two ways a TextGrid could be written: we can either place each utterance in an individual IntervalTier, or each word in its own IntervalTier; we leave that decision up to you. To learn more about TextGrid, visit this page.

# reading; recall we can either interpret each IntervalTier as a word or utterance
tg_utterance = ba.TextGridFile("utterance", path="tg_ut.TextGrid", lang="eng")
tg_word = ba.TextGridFile("word", path="tg_w.TextGrid", lang="eng")

doc1 = tg_utterance.doc
doc2 = tg_word.doc

# writing
tg_utterance = ba.TextGridFile("utterance", doc=doc1)
tg_word = ba.TextGridFile("word", doc=doc2)

tg_utterance.write("tg_ut.TextGrid")
tg_word.write("tg_w.TextGrid")

Questions?

If you have any questions or concerns, please reach out! If something isn't working right, open an issue on GitHub; if you need support, please feel free to email houjun@cmu.edu and macw@cmu.edu.