EchoFlow POC

Hold-to-talk voice transcription with Gemini Pro aggregation. Supports two STT engines:

• Google Chirp 3 — English transcription via Google Cloud Speech-to-Text
• Sarvam AI — Indian language transcription (Hindi, Tamil, Telugu, etc.) with auto language detection

Flow

Hold Left Shift + Left Ctrl → Audio chunked every 5s → STT (parallel)
Release Left Shift + Left Ctrl → Gemini Pro aggregates → Final text output

Prerequisites

• Python 3.10+
• Gemini API key

For Chirp 3 mode:

• Google Cloud project with Speech-to-Text API enabled
• gcloud CLI installed and authenticated

For Sarvam mode:

• Sarvam AI API key (from sarvam.ai)

Setup

1. Install dependencies

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2a. Chirp 3 setup

gcloud auth login
gcloud config set project YOUR_PROJECT_ID
gcloud services enable speech.googleapis.com

export GOOGLE_CLOUD_PROJECT="your-project-id"
export GEMINI_API_KEY="your-gemini-api-key"

# Optional (defaults shown)
export CHIRP_REGION="us-central1"
export CHIRP_RECOGNIZER="_"

2b. Sarvam AI setup

export SARVAM_API_KEY="your-sarvam-api-key"
export GEMINI_API_KEY="your-gemini-api-key"

# Optional — defaults to "unknown" (auto-detect)
export SARVAM_LANGUAGE_CODE="hi-IN"

Run

Via router (edit STT_ENGINE in run.py to switch between "chirp" and "sarvam"):

python run.py

Directly:

python main.py          # Chirp 3
python sarvam_main.py   # Sarvam AI

Usage

1. Hold Left Shift + Left Ctrl to start recording
2. Speak - audio is chunked every 5 seconds and sent to Chirp 3
3. Release Left Shift + Left Ctrl to stop and get aggregated result
4. Paste (Cmd+V) - The final transcribed text is automatically copied to your clipboard. You can paste it anywhere using Cmd + V.
5. Press ESC to exit

Platform Notes

macOS

Grant accessibility permission when prompted: System Preferences → Privacy & Security → Accessibility → Enable for Terminal/IDE

If cursor typing seems flaky, the app now pastes the final text using Cmd+V on macOS, so ensure your target window accepts paste operations and that pbcopy is available (it is installed by default). Set ECHOFLOW_MAC_PASTE=0 to force key-by-key typing instead of the automatic paste on macOS.

Linux

May need to add user to input group:

sudo usermod -aG input $USER
# Log out and back in

Architecture

Single-file per engine. run.py routes between them. No tests, no build step.

• Audio capture: sounddevice.InputStream callback fills buffer
• Chunking: Timer thread drains buffer every 5s
• STT: ThreadPoolExecutor(max_workers=5) sends chunks in parallel
• Aggregation: Gemini aggregates all chunk transcripts into clean text
• Output: Copies to clipboard and pastes via Cmd+V on macOS