ichigo

About

Welcome to Ichigo, a streamlined speech package designed to empower developers with cutting-edge speech models and tools, cultivated by the innovative Ichigo team. The rapidly evolving landscape of speech technology demands a solution that simplifies and unifies speech tasks. Ichigo does just that, enabling developers with straightforward access to powerful models through intuitive Python interfaces or a scalable FastAPI service, leaving behind the tedious intricacies of audio processing so you can focus on what truly matters—deploying and improving your systems.

List of Capabilities

This package does 3 things:

1. Automatic Speech Recognition: Ichigo-ASR
2. Text to Speech: Coming Soon
3. Speech Language Model: Ichigo-LLM (experimental)

It contains only inference code, and caters to most local inference use cases around these three tasks.

Installation

To get started, simply install the package.

pip install ichigo

Ichigo-ASR

Ichigo-ASR is a compact (22M parameters), open-source speech tokenizer for the Whisper-medium model, designed to enhance performance on multilingual with minimal impact on its original English capabilities. Unlike models that output continuous embeddings, Ịchigo-ASR compresses speech into discrete tokens, making it more compatible with large language models (LLMs) for immediate speech understanding. This speech tokenizer has been trained on over ~400 hours of English data and ~1000 hours of Vietnamese data.

Batch Processing

The ichigo package can handle batch processing of audio files using a single line of code, with additional parameters for available for more control.

1. For single files

# Quick one-liner for transcription
from ichigo.asr import transcribe
results = transcribe("path/to/your/file")
# Expected output: "{filename: transcription}"

A transcription.txt will also stored in the same folder as "path/to/your/file"

1. For multiple files (folder)

# Quick one-liner for transcription
from ichigo.asr import transcribe
results = transcribe("path/to/your/folder")
# Expected output: "{filename1: transcription1, filename2: transcription2, ... filenameN: transcriptionN,}"

A subfolder will be created in path/to/your/folder and transcriptions will be stored as filenameN.txt in the subfolder.

API

For integration with frontend, a python fastAPI is also available. This api also does batch processing. Streaming is currently not supported.

1. Start the server

# Uvicorn
cd api && uvicorn asr:app --host 0.0.0.0 --port 8000

# or Docker 
docker compose up -d

1. curl

…

You can also access the API documentation at http://localhost:8000/docs

Ichigo-LLM

:strawberry: Ichigo-LLM is an open, ongoing research experiment to extend a text-based LLM to have native "listening" ability. Think of it as an open data, open weight, on device Siri.

It uses an early fusion technique inspired by Meta's Chameleon paper.

We build train in public:

• Ichigo v0.3 Checkpoint Writeup
• Ichigo v0.2 Checkpoint Writeup
• Ichigo v0.1 Checkpoint Writeup

Ichigo-TTS

Coming Soon

Background

Not your grandfather's speech package

We have modularized the ASR and TTS speech tasks so that they can share components and speak the same language. This is powerful because we can leverage ASR data to train TTS and vice-versa, and using our novel model, Speechless, we can train speech language models without using speech (paper coming soon).

We built this package around what we our vision of the future of speech -- the unification of speech tasks into a single representation framework. Today, many speech models are monolithic models trained end-to-end for a single task. However, subcomponents of these models are in fact reusable for other tasks. This means that ASR fine-tuning can be pre-training for TTS, allowing us to bootstrap better models even with limited training data. We designed this package with the right level of modularity to enable people to train their own models in this manner.

Our package supports underlying abstractions that can support different kinds of models and we welcome researchers to work with us to build these models if you find this way of doing things helpful. You should join our discord community where we can talk about this. A technical report explaining how this framework works and how we think about the future of speech will be uploaded soon.

Everything here is a work in progress, and we welcome all kinds of feedback and collaborations

Benchmarks

Join Us

:strawberry: Ichigo-LLM and Ichigo-ASR is an open research project. We're looking for collaborators, and will likely move towards crowdsourcing speech datasets in the future.

References

…

Acknowledgement

• torchtune: The codebase we built upon
• WhisperSpeech: Text-to-speech model for synthetic audio generation
• llama3: the Family of Models that we based on that has the amazing language capabilities