moonshine-tiny-zh

Usage

Moonshine is supported in Hugging Face 🤗 Transformers. To run the model, first install the Transformers library. For this example, we'll also install 🤗 Datasets to load toy audio dataset from the Hugging Face Hub, and 🤗 Accelerate to reduce the model loading time:

pip install --upgrade pip
pip install --upgrade transformers datasets[audio]

from transformers import MoonshineForConditionalGeneration, AutoProcessor
from datasets import load_dataset, Audio
import torch

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model = MoonshineForConditionalGeneration.from_pretrained('UsefulSensors/moonshine-tiny-zh').to(device).to(torch_dtype)
processor = AutoProcessor.from_pretrained('UsefulSensors/moonshine-tiny-zh')

dataset = load_dataset('UsefulSensors/multilingual_examples', split='zh')
dataset = dataset.cast_column("audio", Audio(processor.feature_extractor.sampling_rate))
sample = dataset[0]["audio"]

inputs = processor(
    sample["array"],
    return_tensors="pt",
    sampling_rate=processor.feature_extractor.sampling_rate
)
inputs = inputs.to(device, torch_dtype)

# to avoid hallucination loops, we limit the maximum length of the generated text based expected number of tokens per second
token_limit_factor = 13 / processor.feature_extractor.sampling_rate
seq_lens = inputs.attention_mask.sum(dim=-1)
max_length = int((seq_lens * token_limit_factor).max().item())

generated_ids = model.generate(**inputs, max_length=max_length)
print(processor.decode(generated_ids[0], skip_special_tokens=True))

Model Details

This Moonshine model is trained for the speech recognition task, capable of transcribing Chinese speech audio into Chinese text. Moonshine AI developed the models to support their business direction of developing real time speech transcription products based on low cost hardware. The following table shows comparisons of common ASR evaluations sets. For more information about evaluation, please refer to the paper.

Release date

September 2025

Model type

Sequence-to-sequence ASR (automatic speech recognition) and speech translation model

Model Use

Evaluated Use

The primary intended users of these models are AI developers that want to deploy Chinese speech recognition systems in platforms that are severely constrained in memory capacity and computational resources. We recognize that once models are released, it is impossible to restrict access to only “intended” uses or to draw reasonable guidelines around what is or is not safe use.

The models are primarily trained and evaluated on Chinese ASR task. They may exhibit additional capabilities, particularly if fine-tuned on certain tasks like voice activity detection, speaker classification, or speaker diarization but have not been robustly evaluated in these areas. We strongly recommend that users perform robust evaluations of the models in a particular context and domain before deploying them.

In particular, we caution against using Moonshine models to transcribe recordings of individuals taken without their consent or purporting to use these models for any kind of subjective classification. We recommend against use in high-risk domains like decision-making contexts, where flaws in accuracy can lead to pronounced flaws in outcomes. The models are intended to transcribe Chinese speech, use of the model for classification is not only not evaluated but also not appropriate, particularly to infer human attributes.

Training Data

The models are trained on 69,800 hours of audio and the corresponding transcripts collected from the internet, as well as datasets openly available and accessible on HuggingFace. The open datasets used are listed in the the accompanying paper.

Performance and Limitations

Our evaluations show that, the models exhibit greater accuracy on standard datasets over existing ASR systems of both similar and larger sizes.

However, like any machine learning model, the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination). We hypothesize that this happens because, given their general knowledge of language, the models combine trying to predict the next word in audio with trying to transcribe the audio itself.

In addition, the sequence-to-sequence architecture of the model makes it prone to generating repetitive texts, which can be mitigated to some degree by beam search and temperature scheduling but not perfectly. It is likely that this behavior and hallucinations may be worse for short audio segments, or segments where parts of words are cut off at the beginning or at the end of the segment.

Broader Implications

We anticipate tha