🎙️ → 📝 → 🗒️ → [🔊🇫🇷] → 🔊
The goal of this tutorial is to build an AI-assisted podcast generator that works across multiple languages. Starting from a recorded draft, we’ll transcribe it, enrich it with an LLM, translate it, and synthesize the result back into audio.
The key idea is to simulate a microservice architecture, where each component runs as its own containerized service. The full pipeline is shown below.
- Pavlos recorded a draft podcast in English, which serves as our starting point.
- The audio file is transcribed using the Google Cloud Speech-to-Text API.
- The resulting text is sent to an LLM to generate an expanded version of the podcast.
- The generated text is synthesized into audio with Google Cloud Text-to-Speech.
- The text is also translated into French (or another language) using Google Translation services.
- The translated text is synthesized into audio again with Google Cloud Text-to-Speech.
- Bonus step: The translated text can also be synthesized with ElevenLabs to recreate Pavlos’ voice.
The pipeline flow is illustrated below:
By completing this tutorial, you’ll gain experience with:
- Containerizing AI/ML workflows step by step.
- Using shared cloud storage (GCS) to connect independent services.
- Securing applications with service account authentication.
Students will form teams for this project. Each team will build the entire pipeline end to end. You won’t just handle one piece—you’ll containerize and connect every component.
All components and their step-by-step instructions are listed below so you can follow along:
- 📝 Task A — transcribe_audio
- 🗒️ Task B — generate_text
- 🔊 Task C — synthesis_audio_en
- 🇫🇷 Task D — translate_text
- 🔊 Task E — synthesis_audio
By the end, every team will have built a complete pipeline that mirrors a real-world microservice architecture: multiple independent services, each containerized, working together to form a larger application.
The overall progress of the mega pipeline can be viewed here.
When building your containers, make sure you update the group name inside your configuration.
This is how we track your progress and display it correctly on the leaderboard.
If you don’t change the group name, your work may overwrite someone else’s, or it won’t be visible under your team.
So please double-check before you push or run your containerized tasks!
In a production pipeline, containerized services talk to each other through APIs, sending requests and responses directly between microservices.
Since we haven’t covered APIs yet, we’ll simplify. Instead of calling one another directly, components will communicate indirectly by writing their outputs to storage, which the next stage will then read as input.
In this tutorial, rather than just using your local disk, components will write to and read from a Google Cloud Storage (GCS) bucket. This shared bucket acts like a common drive for transcripts, generated text, and synthesized audio.
This setup gives you practical hands-on experience now, while preparing you for the API-driven systems we’ll tackle later.
In Google Cloud, a bucket is like a shared online folder where files can be stored and retrieved. Instead of saving outputs locally, our pipeline components will read and write to this shared bucket so all stages can communicate easily.
input_audios/— raw audio files (starting point).text_prompts/— transcripts generated from speech-to-text.text_paragraphs/— expanded text generated by the LLM.text_translated/— translated versions of the text.text_audios/— synthesized audio clips for each paragraph.output_audios/— final audio outputs in French (or another language).output_audios_pp/— French audio outputs in Pavlos’ voice.
The last piece we need in order to access the GCP bucket is authentication. Buckets won’t let you read or write anything unless you are both authenticated (proving who you are) and authorized (having the right permissions).
In this course, you don’t need to authenticate yourself as a person. Instead, you’ll authenticate your app so it can talk to GCP securely. The way we do this is by using a Service Account—part of IAM in GCP.
To keep it simple, you’ll use a JSON credentials file that represents this Service Account. Just download the file and place it inside <app_folder>/secrets/:
Later in the course, we’ll revisit authentication in more depth, but for now, this file is all you need to let your containerized apps talk to the GCP bucket.
🔑 Important Note on Secrets
We do not want to put this JSON file in GitHub — it is a secret, after all. Make sure the secrets/ folder containing the file is not part of your repo. For this tutorial, we’ve already added a .gitignore entry so the file won’t be pushed accidentally. The canonical (best) way to handle this is to keep your secrets folder outside the repo entirely. That’s what we’ll be moving toward later in the course.
Transcribe Audio
python cli.py --download
python cli.py --transcribe
python cli.py --uploadGenerate Text
python cli.py --download
python cli.py --generate
python cli.py --upload
Synthesize Audio (English)
python cli.py --download
python cli.py --synthesis
Translate Text
python cli.py --download
python cli.py --translate
python cli.py --upload
Synthesize Audio (Trnslated)
python cli.py --download
python cli.py --synthesis
- Download from bucket
from google.cloud import storage
# Initiate Storage client
storage_client = storage.Client(project=gcp_project)
# Get reference to bucket
bucket = storage_client.bucket(bucket_name)
# Find all content in a bucket
blobs = bucket.list_blobs(prefix="input_audios/")
for blob in blobs:
print(blob.name)
if not blob.name.endswith("/"):
blob.download_to_filename(blob.name)
- Upload to bucket
from google.cloud import storage
# Initiate Storage client
storage_client = storage.Client(project=gcp_project)
# Get reference to bucket
bucket = storage_client.bucket(bucket_name)
# Destination path in GCS
destination_blob_name = "input_audios/test.mp3"
blob = bucket.blob(destination_blob_name)
blob.upload_from_filename("Path to test.mp3 on local computer")
# Use the official Debian-hosted Python image
FROM python:3.12-slim-bookworm
ARG DEBIAN_PACKAGES="build-essential curl"
# Prevent apt from showing prompts
ENV DEBIAN_FRONTEND=noninteractive
# Python wants UTF-8 locale
ENV LANG=C.UTF-8
# Tell Python to disable buffering so we don't lose any logs.
ENV PYTHONUNBUFFERED=1
# Tell uv to copy packages from the wheel into the site-packages
ENV UV_LINK_MODE=copy
ENV UV_PROJECT_ENVIRONMENT=/home/app/.venv
# This is done for the tutorial only
ENV GOOGLE_APPLICATION_CREDENTIALS=secrets/mega-pipeline.json
# Ensure we have an up to date baseline, install dependencies and
# create a user so we don't run the app as root
RUN set -ex; \
for i in $(seq 1 8); do mkdir -p "/usr/share/man/man${i}"; done && \
apt-get update && \
apt-get upgrade -y && \
apt-get install -y --no-install-recommends $DEBIAN_PACKAGES && \
apt-get clean && \
rm -rf /var/lib/apt/lists/* && \
pip install --no-cache-dir --upgrade pip && \
pip install uv && \
useradd -ms /bin/bash app -d /home/app -u 1000 && \
mkdir -p /app && \
chown app:app /app
# Switch to the new user
USER app
WORKDIR /app
# Copy the source code
COPY --chown=app:app . ./
RUN uv sync
# Entry point
ENTRYPOINT ["/bin/bash"]
# Get into the uv virtual environment shell
CMD ["-c", "source /home/app/.venv/bin/activate && exec bash"]
- Docker Win10 installation - needs WSL2 or Hyper-V enabled: https://docs.docker.com/desktop/windows/install/
- Use
GitBASH to run (which is like a smallerCygwin) - Needed to add pwd in quotes in order to escape the spaces that common in windows directory structures
- Need to prefix docker run with
winptyotherwise I get a "the input device is not a TTY." error winpty docker run --rm -ti --mount type=bind,source="$(pwd)",target=/app generate_text
Solutions to this tutorial can be found here