Simple web service for audio transcription
This is a web service that allows upload of a wav file and returns a transcript in the desired format.
Webscribe uses the nzilbb.ag annotation graph automation framework which includes:
- An abstract object model for modelling speech acts (Annotation Graphs)
- Transcriber implementations that integrate with different ASR systems in order to create annotation graphs representing transcripts of recordings.
- Serializer implementations that transform annotation graphs into different tool formats, e.g. ELAN, Praat TextGrid, WebVTT subtitles, etc.
The default implementation uses WhisperX (based on OpenAI's WhisperX) speech-to-text (STT) system for transcription and optionally diarization, and includes download as ELAN transcript or Praat TextGrid.
These instructions assume installation on a Linux-based server.
Prerequisites include:
- at least version 3.7 of python
(
sudo yum install python3.8 && sudo alternatives --set python3 /usr/bin/python3.8) - python's developer package (
sudo yum install python38-devel.x86_64) - pip3 (
sudo yum install python3-pip) - venv (
sudo yum install python3-all-venv) - ffmpeg (
sudo yum install ffmpeg) - Apache Tomcat (
sudo yum install tomcat)
- Copy webscribe.war to the Tomcat webapps directory.
cp ~/webscribe.war /var/lib/tomcat/webapps/ - Browse to your server's address with
webscribe/appended to the URL (e.g.http://localhost:8080/webscribe/) and upload a short recording.
This will start the WhisperX installation, download the default model, and create the configuration file, which you can edit to change your settings. This can take a few minutes to complete.
The transcriber's configuration is loaded from .../webscribe/transcriber/WhisperXTranscriber/WhisperXTranscriber.cfg
The configuration is text encoded as a URL, e.g.
minShortPauseLength=0.35&minMediumPauseLength=0.7&minLongPauseLength=1.4&shortPauseLabel=%28.%29&mediumPauseLabel=%28..%29&longPauseLabel=%28...%29&maxUtteranceDuration=15.0&utterancePadding=0.5
To change the configuration, you can change the values of parameters in the string,
or add new parameters separated by &. Possible parameters include:
language
: Sets the expected language of the audio (it is otherwise inferred from the recording)
e.g. language=en for speech in English.
diarize
: Enables diarization if set to true - i.e. diarize=true.
In order for diarization to work, you also need to set huggingFaceToken.
huggingFaceToken
: Your huggingface.co authentication token, which is required for downloading
diarization models - e.g. huggingFaceToken=hf_XXXXXXXXXXXXXXXXXXXXX.
See https://huggingface.co/docs/hub/security-tokens for information about
how to generate a token.
minSpeakers
: Minimum number of speakers for diarization, e.g. minSpeakers=2
maxSpeakers
: Maximum number of speakers for diarization, e.g. maxSpeakers=5
model
: Model to use for STT - can be the name of a Whisper model
(e.g. model=large-v3), or the directory containing a custom model
(e.g. model=/opt/whisper/my-adapted-model or model=my-adapted-model
if the model directory is in .../webscribe/transcriber/WhisperXTranscriber/).
The default model is medium)
chunkSize
: The length (in seconds) of audio processed at once. 30
seconds is a balanced default for processing time vs. accuracy.
Smaller chunks can yield more precise boundaries, while larger
chunks can speed up processing. e.g. chunkSize=15
- Click the Browse button and select a .wav file to transcribe.
- Select the transcript format you want.
- Wait patiently while the file uploads to the server (there's a progress bar that indicates how much is uploaded so far).
- Wait patiently while the recording is transcribed (there's a second progress bar that may indicate progress, although generally this stays at 0% for a long time and then is suddenly at 100%). The first time you upload a recording, there is an initial delay while Whisper downloads its models for STT. This is a one-time process, and subsequent recordings don't require this initial delay.
- Save the resulting transcript.