5f087d1335
* config draft * full encoder forward * full decoder forward * fix sdpa and FA2 * fix sdpa and FA2 * moonshine model * moonshine model forward * fix attention with past_key_values * add MoonshineForConditionalGeneration * fix cache handling and causality for cross attention * no causal attention mask for the encoder * model addition (imports etc) * small nit * nits * Update src/transformers/models/moonshine/convert_usefulsensors_to_hf.py Co-authored-by: Joshua Lochner <admin@xenova.com> * add rope_theta * nits * model doc * Update src/transformers/models/auto/configuration_auto.py Co-authored-by: Joshua Lochner <admin@xenova.com> * imports * add MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES * updates modular * make * make fix-copies * ruff check examples fix * fix check_modular_conversion * nit * nits * nits * copied from -> imports * imports fix * integrate attention refacto * modular edge case * remove encoder * convolutions params in config * run modular_model_converter * make * Update docs/source/en/model_doc/moonshine.md Co-authored-by: Joshua Lochner <admin@xenova.com> * MoonshineModelTest * correct typo * make style * integration tests * make * modular convert * name conversion update (up_proj -> fc1 etc) * update config * update MLP * update attention * update encoder layer * update decoder layer * update convolutions parameters * update encoder * remove INPUTS_DOCSTRING * update decoder * update conditional generation * update pretrained model * imports * modular converted * update doc * fix * typo * update doc * update license * update init * split config in file * two classes for MLP * attention from GLM * from GlmRotaryEmbedding * split MLP * apply arthur's review suggestions * apply arthur's review suggestions * apply arthur's review suggestions * auto feature extractor * convert modular * fix + make * convert modular * make * unsplit config * use correct checkpoint * wrap generate * update tests * typos * make * typo * update doc --------- Co-authored-by: Joshua Lochner <admin@xenova.com>
2.7 KiB
Moonshine
Overview
The Moonshine model was proposed in Moonshine: Speech Recognition for Live Transcription and Voice Commands by Nat Jeffries, Evan King, Manjunath Kudlur, Guy Nicholson, James Wang, Pete Warden.
The abstract from the paper is the following:
This paper introduces Moonshine, a family of speech recognition models optimized for live transcription and voice command processing. Moonshine is based on an encoder-decoder transformer architecture and employs Rotary Position Embedding (RoPE) instead of traditional absolute position embeddings. The model is trained on speech segments of various lengths, but without using zero-padding, leading to greater efficiency for the encoder during inference time. When benchmarked against OpenAI's Whisper tiny-en, Moonshine Tiny demonstrates a 5x reduction in compute requirements for transcribing a 10-second speech segment while incurring no increase in word error rates across standard evaluation datasets. These results highlight Moonshine's potential for real-time and resource-constrained applications.
Tips:
- Moonshine improves upon Whisper's architecture:
- It uses SwiGLU activation instead of GELU in the decoder layers
- Most importantly, it replaces absolute position embeddings with Rotary Position Embeddings (RoPE). This allows Moonshine to handle audio inputs of any length, unlike Whisper which is restricted to fixed 30-second windows.
This model was contributed by Eustache Le Bihan (eustlb). The original code can be found here.
Resources
MoonshineConfig
autodoc MoonshineConfig
MoonshineModel
autodoc MoonshineModel - forward - _mask_input_features
MoonshineForConditionalGeneration
autodoc MoonshineForConditionalGeneration - forward - generate