duck_transcriber

Overview

This is a serverless Telegram bot that transcribes voice, audio, and video notes sent to it using the Groq Whisper API. It also stores and retrieves transcriptions using AWS DynamoDB.

How It Works

1. The bot receives a voice, audio, or video note message from a user.
2. It downloads the file from Telegram and checks its duration. If the duration is above a specified limit, it sends a warning message to the user and exits.
3. The bot checks if the transcription already exists in AWS DynamoDB. If it does, the saved transcription is sent back to the user.
4. If no transcription is found, the bot transcribes the audio using the Groq Whisper API.
5. The transcription is sent back to the user as a text message and stored in DynamoDB for future reference.

Supported Commands

• /start: Initializes the bot and provides a welcome message
• /help: Provides information on how to use the bot and its features
• /transcribe: Transcribes the voice, audio, or video note in the reply message
• /translate: Translates (into English) the voice, audio, or video note in the reply message
• /summarize: Summarizes the voice, audio, or video note in the reply message
• /caveman: Transcribes the voice, audio, or video note in the reply message in a "caveman" style

Developer Commands

• /check: Report Telegram webhook queue status (pending updates, last error)
• /reset: Reset webhook with drop_pending_updates=true

These commands are only available to the developer whose Telegram user ID is set via the DEV_TELEGRAM_ID environment variable. If DEV_TELEGRAM_ID is not set, the bot ignores these commands.

Why this exists: when the Lambda handler returns a non-200 response, Telegram will retry the same webhook delivery. This can create a resend loop and queue backlog. The developer commands allow monitoring the queue and forcefully resetting the webhook with drop_pending_updates=true to clear the backlog.

Maybe in the future, I'll make a fork of Teloxide that would be specially made to be run on serverless functions.

Technical Details

• The bot is built using the teloxide crate for interacting with the Telegram API.
• The transcription is done using the reqwest crate to send a request to the Groq Whisper API.
• The file is first converted to a 16 kHz mono FLAC file using FFmpeg before being sent to the Groq Whisper API.
• The bot uses AWS DynamoDB to store and retrieve transcriptions, ensuring that repeated requests for the same audio do not require retranscription.
• The bot is deployed as a serverless function using AWS Lambda.

Environment Variables

• TELEGRAM_BOT_TOKEN: the token for the Telegram bot.
• GROQ_API_KEY: the API key for the Groq Whisper API.
• DYNAMODB_TABLE: the name of the DynamoDB table where transcriptions are stored.
• DEV_TELEGRAM_ID: Telegram user ID allowed to run developer commands (/check, /reset). If unset, the commands are ignored.
• LAMBDA_URL: The URL of the Lambda function. This is used to set the webhook in the Telegram API when resetting it.

Deployment

Before deploying this bot, ensure you have the following prerequisites installed:

• AWS CLI: Follow the instructions here to install the AWS Command Line Interface.
• cargo-lambda: Install cargo-lambda (not with cargo, it doesn't support cross compilation) by following the instructions here.

Build and Deploy

To build:

cargo lambda build --release --arm64

To deploy:

cargo lambda deploy

Create FFmpeg Lambda Layer

1. Download FFmpeg Static Build: Download the ARM64 static build of FFmpeg for Amazon Linux 2023:

   wget https://johnvansickle.com/ffmpeg/releases/ffmpeg-release-arm64-static.tar.xz
   tar -xf ffmpeg-release-arm64-static.tar.xz
   mv ffmpeg-*-arm64-static/ffmpeg ./ffmpeg
   chmod +x ffmpeg

2. Create the Layer Directory: Create a directory structure for the Lambda layer:

   mkdir -p ffmpeg-layer/bin
   mv ffmpeg ffmpeg-layer/bin/

3. Zip the Layer: Compress the ffmpeg-layer directory:

   zip -r ffmpeg-layer.zip ffmpeg-layer

4. Publish the Layer: Publish the layer to AWS Lambda using the AWS CLI:

   aws lambda publish-layer-version \
     --layer-name ffmpeg-arm64 \
     --description "FFmpeg static binary for ARM64 on Amazon Linux 2023" \
     --zip-file fileb://ffmpeg-layer.zip \
     --compatible-runtimes provided.al2023 \
     --compatible-architectures arm64

5. Attach the Layer to Your Lambda Function: Update the Cargo.toml file to include the ARN of the FFmpeg Lambda layer

AWS Lambda Permissions

Ensure that your AWS Lambda function has the necessary permissions to access DynamoDB. You will need to attach a policy that grants the Lambda function read and write permissions to the DynamoDB table. This can be done by attaching the AWSLambdaDynamoDBExecutionRole managed policy or by creating a custom policy with the necessary permissions.

License

Do literally whatever you want with this code. I don't care.

Contributing

Contributions are welcome! If you'd like to help improve this bot, please open a pull request with your changes.