usefulsensors / moonshine

Fast and accurate automatic speech recognition (ASR) for edge devices
MIT License
2.22k stars 98 forks source link

Moonshine

[Blog] [Paper] [Model Card] [Podcast]

Moonshine is a family of speech-to-text models optimized for fast and accurate automatic speech recognition (ASR) on resource-constrained devices. It is well-suited to real-time, on-device applications like live transcription and voice command recognition. Moonshine obtains word-error rates (WER) better than similarly-sized tiny.en and base.en Whisper models from OpenAI on the datasets used in the OpenASR leaderboard maintained by HuggingFace:

TinyBase
| WER | Moonshine | Whisper | | ---------- | --------- | ------- | | Average | **12.66** | 12.81 | | AMI | 22.77 | 24.24 | | Earnings22 | 21.25 | 19.12 | | Gigaspeech | 14.41 | 14.08 | | LS Clean | 4.52 | 5.66 | | LS Other | 11.71 | 15.45 | | SPGISpeech | 7.70 | 5.93 | | Tedlium | 5.64 | 5.97 | | Voxpopuli | 13.27 | 12.00 | | WER | Moonshine | Whisper | | ---------- | --------- | ------- | | Average | **10.07** | 10.32 | | AMI | 17.79 | 21.13 | | Earnings22 | 17.65 | 15.09 | | Gigaspeech | 12.19 | 12.83 | | LS Clean | 3.23 | 4.25 | | LS Other | 8.18 | 10.35 | | SPGISpeech | 5.46 | 4.26 | | Tedlium | 5.22 | 4.87 | | Voxpopuli | 10.81 | 9.76 |

Moonshine's compute requirements scale with the length of input audio. This means that shorter input audio is processed faster, unlike existing Whisper models that process everything as 30-second chunks. To give you an idea of the benefits: Moonshine processes 10-second audio segments 5x faster than Whisper while maintaining the same (or better!) WER.

Moonshine Base is approximately 400MB, while Tiny is around 190MB. Both publicly-released models currently support English only.

This repo hosts inference code and demos for Moonshine.

Installation

We like uv for managing Python environments, so we use it here. If you don't want to use it, simply skip the first step and leave uv off of your shell commands.

1. Create a virtual environment

First, install uv for Python environment management.

Then create and activate a virtual environment:

uv venv env_moonshine
source env_moonshine/bin/activate

2. Install the Moonshine package

The moonshine inference code is written in Keras and can run with each of the backends that Keras supports: Torch, TensorFlow, and JAX. The backend you choose will determine which flavor of the moonshine package to install. If you're just getting started, we suggest installing the (default) Torch backend:

uv pip install useful-moonshine@git+https://github.com/usefulsensors/moonshine.git

To run the provided inference code, you have to instruct Keras to use the PyTorch backend by setting an environment variable:

export KERAS_BACKEND=torch

To run with the TensorFlow backend, run the following to install Moonshine and set the environment variable:

uv pip install useful-moonshine[tensorflow]@git+https://github.com/usefulsensors/moonshine.git
export KERAS_BACKEND=tensorflow

To run with the JAX backend, run the following:

uv pip install useful-moonshine[jax]@git+https://github.com/usefulsensors/moonshine.git
export KERAS_BACKEND=jax
# Use useful-moonshine[jax-cuda] for jax on GPU

To run with ONNX runtime that is supported on platforms, run the following:

uv pip install useful-moonshine[onnx]@git+https://github.com/usefulsensors/moonshine.git

3. Try it out

You can test Moonshine by transcribing the provided example audio file with the .transcribe function:

python
>>> import moonshine
>>> moonshine.transcribe(moonshine.ASSETS_DIR / 'beckett.wav', 'moonshine/tiny')
['Ever tried ever failed, no matter try again, fail again, fail better.']

The first argument is a path to an audio file and the second is the name of a Moonshine model. moonshine/tiny and moonshine/base are the currently available models. Use the moonshine.transcribe_with_onnx function to use the ONNX runtime for inference. The parameters are the same as they are for moonshine.transcribe.

Examples

The Moonshine models can be used with a variety of different runtimes and applications, so we've included code samples showing how to use them in different situations. The moonshine/demo folder in this repository also has more information on many of them.

Onnx Standalone

The latest versions of the Onnx Moonshine models are available on HuggingFace at huggingface.co/UsefulSensors/moonshine/tree/main/onnx. You can find an example Python script and more information about running them in the demo folder.

Live Captions

You can try the Moonshine models with live input from a microphone on many platforms with the live captions demo.

Running in the Browser

You can try out the Moonshine models on your device in a web browser with our HuggingFace space. We've included the source for this demo in this repository; this is a great starting place for those wishing to build web-based applications with Moonshine.

CTranslate2

The files for the CTranslate2 versions of Moonshine are available at huggingface.co/UsefulSensors/moonshine/tree/main/ctranslate2, but they require a pull request to be merged before they can be used with the mainline version of the framework. Until then, you should be able to try them with our branch, with this example script.

HuggingFace Transformers

Both models are also available on the HuggingFace hub and can be used with the transformers library, as follows:

from transformers import AutoModelForSpeechSeq2Seq, AutoConfig, PreTrainedTokenizerFast

import torchaudio
import sys

audio, sr = torchaudio.load(sys.argv[1])
if sr != 16000:
  audio = torchaudio.functional.resample(audio, sr, 16000)

# 'usefulsensors/moonshine-base' for the base model
model = AutoModelForSpeechSeq2Seq.from_pretrained('usefulsensors/moonshine-tiny', trust_remote_code=True)
tokenizer = PreTrainedTokenizerFast.from_pretrained('usefulsensors/moonshine-tiny')

tokens = model(audio)
print(tokenizer.decode(tokens[0], skip_special_tokens=True))

TODO

Known Issues

UserWarning: You are using a softmax over axis 3 of a tensor of shape torch.Size([1, 8, 1, 1])

This is a benign warning arising from Keras. For the first token in the decoding loop, the attention score matrix's shape is 1x1, which triggers this warning. You can safely ignore it, or run with python -W ignore to suppress the warning.

Citation

If you benefit from our work, please cite us:

@misc{jeffries2024moonshinespeechrecognitionlive,
      title={Moonshine: Speech Recognition for Live Transcription and Voice Commands}, 
      author={Nat Jeffries and Evan King and Manjunath Kudlur and Guy Nicholson and James Wang and Pete Warden},
      year={2024},
      eprint={2410.15608},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2410.15608}, 
}