ivrit

Python package providing wrappers around ivrit.ai's capabilities.

Installation

pip install ivrit

Usage

Audio Transcription

The ivrit package provides audio transcription functionality using multiple engines.

Basic Usage

import ivrit

# Transcribe a local audio file
model = ivrit.load_model(engine="faster-whisper", model="ivrit-ai/whisper-large-v3-turbo-ct2")
result = model.transcribe(path="audio.mp3")

# With custom device
model = ivrit.load_model(engine="faster-whisper", model="ivrit-ai/whisper-large-v3-turbo-ct2", device="cpu")
result = model.transcribe(path="audio.mp3")

print(result["text"])

Transcribe from URL

# Transcribe audio from a URL
model = ivrit.load_model(engine="faster-whisper", model="ivrit-ai/whisper-large-v3-turbo-ct2")
result = model.transcribe(url="https://example.com/audio.mp3")

print(result["text"])

Streaming Results

# Get results as a stream (generator)
model = ivrit.load_model(engine="faster-whisper", model="base")
for segment in model.transcribe(path="audio.mp3", stream=True, verbose=True):
    print(f"{segment.start:.2f}s - {segment.end:.2f}s: {segment.text}")

# Or use the model directly
model = ivrit.FasterWhisperModel(model="base")
for segment in model.transcribe(path="audio.mp3", stream=True):
    print(f"{segment.start:.2f}s - {segment.end:.2f}s: {segment.text}")

# Access word-level timing
for segment in model.transcribe(path="audio.mp3", stream=True):
    print(f"Segment: {segment.text}")
    for word in segment.extra_data.get('words', []):
        print(f"  {word['start']:.2f}s - {word['end']:.2f}s: '{word['word']}'")

Async Transcription (RunPod Only)

For RunPod models, you can use async transcription for better performance:

import asyncio
from ivrit.audio import load_model

async def transcribe_async():
    # Load RunPod model
    model = load_model(
        engine="runpod",
        model="large-v3-turbo",
        api_key="your-api-key",
        endpoint_id="your-endpoint-id"
    )
    
    # Stream results asynchronously
    async for segment in model.transcribe_async(path="audio.mp3", language="he"):
        print(f"{segment.start:.2f}s - {segment.end:.2f}s: {segment.text}")

# Run the async function
asyncio.run(transcribe_async())

Note: Async transcription is only available for RunPod models. The sync transcribe() method uses the original sync implementation.

API Reference

load_model()

Load a transcription model for the specified engine and model.

Parameters

• engine (str): Transcription engine to use. Options: "faster-whisper", "stable-ts"
• model (str): Model name for the selected engine
• device (str, optional): Device to use for inference. Default: "auto". Options: "auto", "cpu", "cuda", "cuda:0", etc.
• model_path (str, optional): Custom path to the model (for faster-whisper)

Returns

• TranscriptionModel object that can be used for transcription

Raises

• ValueError: If the engine is not supported
• ImportError: If required dependencies are not installed

transcribe() and transcribe_async()

Transcribe audio using the loaded model.

Parameters

• path (str, optional): Path to the audio file to transcribe
• url (str, optional): URL to download and transcribe
• blob (str, optional): Base64 encoded blob data to transcribe
• language (str, optional): Language code for transcription (e.g., 'he' for Hebrew, 'en' for English)
• stream (bool, optional): Whether to return results as a generator (True) or full result (False) - only for transcribe()
• diarize (bool, optional): Whether to enable speaker diarization
• verbose (bool, optional): Whether to enable verbose output
• **kwargs: Additional keyword arguments for the transcription model

Returns

• transcribe(): If stream=True: Generator yielding transcription segments, If stream=False: Complete transcription result as dictionary
• transcribe_async(): AsyncGenerator yielding transcription segments

Raises

• ValueError: If multiple input sources are provided, or none is provided
• FileNotFoundError: If the specified path doesn't exist
• Exception: For other transcription errors

Note: transcribe_async() is only available for RunPod models and always returns an AsyncGenerator.

Architecture

The ivrit package uses an object-oriented design with a base TranscriptionModel class and specific implementations for each transcription engine.

Model Classes

• TranscriptionModel: Abstract base class for all transcription models
• FasterWhisperModel: Implementation for the Faster Whisper engine

Usage Patterns

Pattern 1: Using load_model() (Recommended)

# Step 1: Load the model
model = ivrit.load_model(engine="faster-whisper", model="base")

# Step 2: Transcribe audio
result = model.transcribe(path="audio.mp3")

Pattern 2: Direct Model Creation

# Create model directly
model = ivrit.FasterWhisperModel(model="base")

# Use the model
result = model.transcribe(path="audio.mp3")

Multiple Transcriptions

For multiple transcriptions, load the model once and reuse it:

# Load model once
model = ivrit.load_model(engine="faster-whisper", model="base")

# Use for multiple transcriptions
result1 = model.transcribe(path="audio1.mp3")
result2 = model.transcribe(path="audio2.mp3")
result3 = model.transcribe(path="audio3.mp3")

Installation

Basic Installation

pip install ivrit

With Faster Whisper Support

pip install ivrit[faster-whisper]

Supported Engines

faster-whisper

Fast and accurate speech recognition using the Faster Whisper model.

Model Class: FasterWhisperModel

Available Models: base, large, small, medium, large-v2, large-v3

Features:

• Word-level timing information
• Language detection with confidence scores
• Support for custom devices (CPU, CUDA, etc.)
• Support for custom model paths
• Streaming transcription

Dependencies: faster-whisper>=1.1.1

stable-ts

Stable and reliable transcription using Stable-TS models.

Status: Not yet implemented

Development

Installation for Development

git clone <repository-url>
cd ivrit
pip install -e ".[dev]"

Running Tests

pytest

Code Formatting

black .
isort .

Bounty rules

Like our bounties, and want to help? Here's how this works:

1. You pick a bounty you're interested in, and let us know. We discuss it together to make sure you understand the issue.
2. You let us know you're on it; we lock it for you for 2 weeks so you can develop, review and merge your code.
3. The ONLY metric for whether you met the bounty goal is whether we decide to merge your PR. Our key focus with reviews is to ensure high code and product quality.
4. Once your PR is merged, you receive the bounty award.

You can use any tool you'd like to write your code, including AI. Note that during review you will be asked questions about the code; if you are unable to explain what it does, or how (sometimes the case when doing Vibe coding), your PR will be discarded and you will not be able to reapply for this issue.

Reviews may be done live.

License

MIT License - see LICENSE file for details.