Article
Extremely Fast and Accurate Transcription with Granite Speech 5.0 Turbo CTC
ibm-granite
β’ β’ 33
k0 c@n gj@j th!ch π
mp1704
AI & ML interests
None yet
Recent Activity
upvoted an article 19 days ago
Extremely Fast and Accurate Transcription with Granite Speech 5.0 Turbo CTC liked a model about 1 month ago
allenai/GraspMolmo liked a dataset about 2 months ago
bones-studio/seedOrganizations
upvoted an article 19 days ago
liked a
model about 1 month ago
liked a
dataset about 2 months ago
upvoted a paper 2 months ago
liked 2
Spaces 2 months ago
liked a
model 3 months ago
updated a
collection 3 months ago
liked a
model 3 months ago
upvoted 5
articles 3 months ago
Article
Mastering Tensor Dimensions in Transformers
not-lain
β’ β’ 204 Article
Visualizing How VLMs Work
not-lain
β’ β’ 58 Article
KV Caching Explained: Optimizing Transformer Inference Efficiency
not-lain
β’ β’ 413 Article
Recipe: Preparing Multilingual Speech Datasets for TTS Training
PHBJT
β’ β’ 21 Article
From Zero to GPU: A Guide to Building and Scaling Production-Ready CUDA Kernels
drbh, danieldk
β’ β’ 111 liked a
model 3 months ago
reacted to matteospanio's post with π 3 months ago
Post
7099
πΆ Released mule-torch β an unofficial PyTorch port of MULE (SF-NFNet-F0), SiriusXM/Pandora's music-audio embedding model (McCallum et al., ISMIR 2022).
No retraining: I re-implemented the architecture in pure PyTorch and transferred the original TensorFlow weights, then checked it layer by layer against the genuine TF pipeline.
β End-to-end clip-embedding cosine 0.9999999 vs the original
β ONNX backbone parity < 1e-6
β 62.35M params (paper: ~62.4M)
β Batched, GPU-native, ONNX-exportable β none of which the original
π€ Weights: matteospanio/mule
π» Code: https://github.com/matteospanio/mule-torch
π¦ PyPI: https://pypi.org/project/mule-torch/
The fun bug: parity was perfect through every conv but the block output was anti-correlated (cos = β1). Cause: the learnable skip-init gains couldn't be mapped by layer name (Keras scrambles the order) β they had to be recovered from the graph.
β οΈ Unofficial, community port β not affiliated with or endorsed by the original authors. All credit to them; please cite the paper. Weights inherit CC-BY-NC-4.0.
No retraining: I re-implemented the architecture in pure PyTorch and transferred the original TensorFlow weights, then checked it layer by layer against the genuine TF pipeline.
β End-to-end clip-embedding cosine 0.9999999 vs the original
β ONNX backbone parity < 1e-6
β 62.35M params (paper: ~62.4M)
β Batched, GPU-native, ONNX-exportable β none of which the original
Analysis pipeline doespip install mule-torchfrom mule_torch import MuleModel
emb = MuleModel.from_pretrained()(waveform) # (B, T)@16kHz -> (B, 1728)π€ Weights: matteospanio/mule
π» Code: https://github.com/matteospanio/mule-torch
π¦ PyPI: https://pypi.org/project/mule-torch/
The fun bug: parity was perfect through every conv but the block output was anti-correlated (cos = β1). Cause: the learnable skip-init gains couldn't be mapped by layer name (Keras scrambles the order) β they had to be recovered from the graph.
β οΈ Unofficial, community port β not affiliated with or endorsed by the original authors. All credit to them; please cite the paper. Weights inherit CC-BY-NC-4.0.
updated a
dataset 3 months ago
published a
dataset 3 months ago
liked a
model 3 months ago
upvoted an article 4 months ago
Article
Building Conversational AI: A Deep Dive into Voice Agent Architectures and Best Practices
abdeljalilELmajjodi
β’ β’ 28