Improve model card: Update paper link, add GitHub, abstract, and new tags

README.md

@@ -1,57 +1,49 @@
 ---
-license: mit
 language:
 - en
 library_name: transformers
+license: mit
 pipeline_tag: automatic-speech-recognition
-arxiv: https://arxiv.org/abs/2410.15608
+paper: https://huggingface.co/papers/2509.02523
+tags:
+- tiny-asr
+- edge-ai
+- monolingual
+- speech-to-text
 ---
-# Moonshine
 
-[[Blog]](https://petewarden.com/2024/10/21/introducing-moonshine-the-new-state-of-the-art-for-speech-to-text/) [[Paper]](https://arxiv.org/abs/2410.15608) [[Installation]](https://github.com/usefulsensors/moonshine/blob/main/README.md) [[Podcast]](https://notebooklm.google.com/notebook/d787d6c2-7d7b-478c-b7d5-a0be4c74ae19/audio)
+# Flavors of Moonshine: Tiny Specialized ASR Models for Edge Devices
+
+[[Blog]](https://petewarden.com/2024/10/21/introducing-moonshine-the-new-state-of-the-art-for-speech-to-text/) [[Paper]](https://huggingface.co/papers/2509.02523) [[Code]](https://github.com/moonshine-ai/moonshine) [[Podcast]](https://notebooklm.google.com/notebook/d787d6c2-7d7b-478c-b7d5-a0be4c74ae19/audio)
 
 This is the model card for running the automatic speech recognition (ASR) models (Moonshine models) trained and released by Useful Sensors.
 
-Following [Model Cards for Model Reporting (Mitchell et al.)](https://arxiv.org/abs/1810.03993), we're providing some information about the automatic speech recognition model. More information on how these models were trained and evaluated can be found [in the paper](https://arxiv.org/abs/2410.15608). Note, a lot of the text has been copied verbatim from the [model card](https://github.com/openai/whisper/blob/main/model-card.md) for the Whisper model developed by OpenAI, because both models serve identical purposes, and carry identical risks.
+Following [Model Cards for Model Reporting (Mitchell et al.)](https://arxiv.org/abs/1810.03993), we're providing some information about the automatic speech recognition model. More information on how these models were trained and evaluated can be found [in the paper](https://huggingface.co/papers/2509.02523). Note, a lot of the text has been copied verbatim from the [model card](https://github.com/openai/whisper/blob/main/model-card.md) for the Whisper model developed by OpenAI, because both models serve identical purposes, and carry identical risks.
 
-## Usage
+## Abstract
+We present the Flavors of Moonshine, a suite of tiny automatic speech recognition (ASR) models specialized for a range of underrepresented languages. Prevailing wisdom suggests that multilingual ASR models outperform monolingual counterparts by exploiting cross-lingual phonetic similarities. We challenge this assumption, showing that for sufficiently small models (27M parameters), training monolingual systems on a carefully balanced mix of high-quality human-labeled, pseudo-labeled, and synthetic data yields substantially superior performance. On average, our models achieve error rates 48% lower than the comparably sized Whisper Tiny model, outperform the 9x larger Whisper Small model, and in most cases match or outperform the 28x larger Whisper Medium model. These results advance the state of the art for models of this size, enabling accurate on-device ASR for languages that previously had limited support. We release Arabic, Chinese, Japanese, Korean, Ukrainian, and Vietnamese Moonshine models under a permissive open-source license.
 
-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:
+## Usage
 
-```bash
-pip install --upgrade pip
-pip install --upgrade transformers datasets[audio]
-```
+Moonshine models are available on the Hugging Face hub and can be used with the `transformers` library, as follows:
 
 ```python
-from transformers import MoonshineForConditionalGeneration, AutoProcessor
-from datasets import load_dataset, Audio
 import torch
+from transformers import AutoProcessor, MoonshineForConditionalGeneration
+from datasets import load_dataset
 
-device = "cuda:0" if torch.cuda.is_available() else "cpu"
-torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+processor = AutoProcessor.from_pretrained("UsefulSensors/moonshine-tiny")
+model = MoonshineForConditionalGeneration.from_pretrained("UsefulSensors/moonshine-tiny")
 
-model = MoonshineForConditionalGeneration.from_pretrained('UsefulSensors/moonshine-tiny').to(device).to(torch_dtype)
-processor = AutoProcessor.from_pretrained('UsefulSensors/moonshine-tiny')
+ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+audio_array = ds[0]["audio"]["array"]
 
-dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
-dataset = dataset.cast_column("audio", Audio(processor.feature_extractor.sampling_rate))
-sample = dataset[0]["audio"]
+inputs = processor(audio_array, return_tensors="pt")
 
-inputs = processor(
-    sample["array"], 
-    return_tensors="pt",
-    sampling_rate=processor.feature_extractor.sampling_rate
-)
-inputs = inputs.to(device, torch_dtype)
+generated_ids = model.generate(**inputs)
 
-# to avoid hallucination loops, we limit the maximum length of the generated text based expected number of tokens per second
-token_limit_factor = 6.5 / processor.feature_extractor.sampling_rate  # Maximum of 6.5 tokens per second
-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))
+transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+print(transcription)
 ```
 
 ## Model Details
@@ -73,7 +65,7 @@ Sequence-to-sequence ASR (automatic speech recognition) and speech translation m
 
 ### Paper & samples
 
-[Paper](https://arxiv.org/abs/2410.15608) / [Blog](https://petewarden.com/2024/10/21/introducing-moonshine-the-new-state-of-the-art-for-speech-to-text/)
+[Paper](https://huggingface.co/papers/2509.02523) / [Blog](https://petewarden.com/2024/10/21/introducing-moonshine-the-new-state-of-the-art-for-speech-to-text/)
 
 ## Model Use
 
@@ -87,7 +79,7 @@ In particular, we caution against using Moonshine models to transcribe recording
 
 ## Training Data
 
-The models are trained on 200,000 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](https://arxiv.org/abs/2410.15608).
+The models are trained on 200,000 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 accompanying paper](https://huggingface.co/papers/2509.02523).
 
 ## Performance and Limitations
 
@@ -95,7 +87,7 @@ Our evaluations show that, the models exhibit greater accuracy on standard datas
 
 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 the end of the segment.
+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
 
@@ -106,13 +98,13 @@ There are also potential dual-use concerns that come with releasing Moonshine. W
 ## Citation
 If you benefit from our work, please cite us:
 ```
-@misc{jeffries2024moonshinespeechrecognitionlive,
-      title={Moonshine: Speech Recognition for Live Transcription and Voice Commands}, 
+@misc{jeffries2025flavorsmoonshine,
+      title={Flavors of Moonshine: Tiny Specialized ASR Models for Edge Devices}, 
       author={Nat Jeffries and Evan King and Manjunath Kudlur and Guy Nicholson and James Wang and Pete Warden},
-      year={2024},
-      eprint={2410.15608},
+      year={2025},
+      eprint={2509.02523},
       archivePrefix={arXiv},
       primaryClass={cs.SD},
-      url={https://arxiv.org/abs/2410.15608}, 
+      url={https://arxiv.org/abs/2509.02523}, 
 }
-```
+```