Python script to update submissions of a form with an AI-generated transcript of its audio attachment(s)

I've been playing a bit with OpenAI's Whisper transcription library. The open source library is relatively small and can be run on simple hardware. It's highly accurate for English even with relatively small models. It can also do translation and there's good information on how likely it is to be correct for each language on the Github page.

I put together a small script that can be used to update submissions of a form with a transcript of an audio file from that submission. This approach would be useful in contexts where it's helpful to have the transcript in Central so that it can be part of the submission review process, for example. If you don't use Central as your primary source of truth, you could do this transcription after downloading your full dataset (as part of your Extract-Transform-Load pipeline).

For projects with relatively short audio files and low submission volume (<1000 per day), it should be fine to run this on the same host as Central with a cron job that runs outside of data collection hours. For large submission volume it may be more cost effective and robust to use a paid API (e.g. OpenAI's).

I used a form with an audio field and a text field for the transcript. I don't want data collectors to see the text field so I included a hidden field that makes the transcript non-relevant. The update script injects the transcript in the text field and changes the value of the hidden field so that the transcript is visible when viewing or editing the submission. I left the text field editable so that a person could further edit the transcript. It could alternatively be made read-only. The form: recordings.xlsx (9.6 KB)

The script is not extensively tested and doesn't do robust error handling. I'm happy to answer any questions and continue to iterate on it if there's interest. Feel free to take the ideas and make it your own!

CC @chrissyhroberts

import logging
import os
import uuid
import whisper

from logging.handlers import RotatingFileHandler
from pyodk.client import Client
from xml.etree import ElementTree as ET

FORM_ID = "recordings"
LOG = "transcriber.log"

CLIENT = Client()
LOGGER = logging.getLogger("transcriber")
TRANSCRIPTION_MODEL = whisper.load_model("base.en")

def configure_logger(filename):
    LOGGER.setLevel(logging.INFO)
    handler = RotatingFileHandler(filename, maxBytes=5 * 1024 * 1024, backupCount=2)
    formatter = logging.Formatter(fmt='%(asctime)s %(message)s', datefmt='%m/%d/%Y %I:%M:%S %p')
    handler.setFormatter(formatter)
    LOGGER.addHandler(handler)

    logging.basicConfig(level=logging.INFO, format='%(asctime)s %(message)s', datefmt='%m/%d/%Y %I:%M:%S %p')

configure_logger(LOG)
LOGGER.info("##### Running #####")

# Download submissions in the 'received' review state (not yet updated by this script)
new_submissions = CLIENT.submissions.get_table(form_id="recordings", filter=f"__system/reviewState eq null")["value"]
new_submissions = {sub['__id']: sub['audio'] for sub in new_submissions}

for instanceid in new_submissions:
    audio_bytes = CLIENT.get(f'/projects/{CLIENT.config.central.default_project_id}/forms/{FORM_ID}/submissions/{instanceid}/attachments/{new_submissions[instanceid]}')
    try:
        with open(new_submissions[instanceid], 'wb') as audio_file:
            audio_file.write(audio_bytes.content)
        
        LOGGER.info(f"Transcribing {instanceid}")
        transcript = TRANSCRIPTION_MODEL.transcribe(new_submissions[instanceid])['text']
    finally:
        os.remove(new_submissions[instanceid])

    submission = CLIENT.get(f'/projects/{CLIENT.config.central.default_project_id}/forms/{FORM_ID}/submissions/{instanceid}.xml').text
    root = ET.fromstring(submission)
    root.find('show_transcript').text = 'y'
    transcript_node = ET.SubElement(root, 'transcript')
    transcript_node.text = transcript

    meta_node = root.find('meta')
    instanceid_node = meta_node.find('instanceID')
    deprecatedid_node = meta_node.find('deprecatedID')
    if deprecatedid_node is None:
        deprecatedid_node = ET.SubElement(meta_node, 'deprecatedID')
    deprecatedid_node.text = instanceid_node.text
    instanceid_node.text = 'uuid:' + str(uuid.uuid4())

    try:
        xml = ET.tostring(root)
    except Exception as e:
        LOGGER.error(e)
    else:
        CLIENT.submissions.edit(instanceid, xml, form_id=FORM_ID, comment="Transcribed by Whisper v20230314 with base.en")

if len(new_submissions) == 0:
    LOGGER.info("No new submissions")
7 Likes

I love this!

We've been working with Whisper AI for a while and are generally very impressed with it.
It seems to work really well for English, French, Spanish etc, and the automatic translations are also remarkably good! I can imagine a version of this script which looks for a hidden variable / flag in the dataset/form which includes a target language and triggers the script to automatically update the form both with the original language transcript and a translated version.

A lot of our interest in Whisper AI has been for languages such as Kiswahili (it has a great range of languages, which is more than can be said for Zoom and other tools' built in transcriptions). Our work with it so far indicates that like in the early days of automatic transcriptions for English, there's some distance left to travel with the less globally ubiquitous languages because the libraries are pretty small. Having said that, the transcriptions are still a good framework for the sometimes laborious task of manual transcription. There's obviously huge value in developing functions like this for ODK users.

I know this is alpha, but I can see this being of huge value to some but not all people collecting audio files. I wonder if a whisper module could run in the background of the core Central Docker release (i.e. not as a standalone script), and could trigger on the basis of a flag in the appearance column for audio type variables.

1 Like

If anyone has a project where'd they'd like to try this capability, email me at yanokwa@getodk.org. We can add this capability on ODK Cloud and use OpenAI's paid API so it works with large amounts of data and you don't have to install any scripts.