onnx-asr v0.11.0

Highlights

• Improved preprocessor performance: added a NumPy version for CPU (up to 10x speedup) and updated ONNX versions with DirectML/WebGPU support.
• Updated Silero VAD to current version and added PyAnnote VAD implementation.
• Updated benchmarks, comparisons and installation guide in docs.
• Removed PyTorch from build time dependencies.
• Added Canary 1B Flash model.

What's Changed

• Refactoring and updating model loader by @istupakov in #71, #72, #74
• Updated preprocessors, added NumPy preprocessors by @istupakov in #81, #92
• Updated deps and CI by @istupakov in #70, #76, #77, #89, #90, #97, #100, #103
• Add WeSpeaker embeddings model by @istupakov in #86
• Move content from Readme to Docs by @istupakov in #91
• Update benchmarks and comparison with original models by @istupakov in #85, #98, #105
• Update Silero VAD to v6.2 by @istupakov in #107
• Add PyAnnote VAD implementation by @FTWsGit in #96
• Fix missed logprobs in VAD recognize_batch by @istupakov in #102
• Fix Nemo Canary model decoding by @istupakov in #106
• Other small fixes by @istupakov in #83, #87, #94, #108

New Contributors

• @FTWsGit made their first contribution in #96

Full Changelog: v0.10.2...v0.11.0

onnx-asr v0.10.2

What's Changed

• Update build, CI, tests and dependencies by @istupakov in #66, #67, #69
• Add Docs via Material for MkDocs and GitHub Pages, update docstrings, Readme and CI by @istupakov in #68

Full Changelog: v0.10.1...v0.10.2

onnx-asr v0.10.1

What's Changed

• Add workaround for TensorRT fp16 models by @istupakov in #61
• Improve typing, CLI and readme by @istupakov in #62
• Add Coveralls to CI by @istupakov in #63

Full Changelog: v0.10.0...v0.10.1

onnx-asr v0.10.0

Highlights

• Improved TensorRT support: 10x speedup on GigaAM v2/v3 CTC and 3x speedup on Parakeet TDT v2/v3!
• Optimized the default ONNX configuration to automatically exclude unsupported providers.
• Installation from source has been simplified - you can now just use pip / uv.
• You can specify the folder where to download the model.
• The cpu_preprocessing option is deprecated and ignored, use preprocessor_config and resampler_config instead.

What's Changed

• Fixed concurrent preprocessor on numpy 1.26 by @istupakov in #49
• Improved models downloading by @istupakov in #51, #52
• Update CI and tests, move from PDM to uv+hatchling by @istupakov in #30, #58
• Add TensorRT support and benchmarks by @istupakov in #55
• Added logprobs to recognize results by @istupakov in #54
• Optimize config selection by @istupakov in #59
• Update documentaion by @istupakov in #57

Full Changelog: v0.9.1...v0.10.0

onnx-asr v0.9.1

What's Changed

• Added separate ONNX options for preprocessor and resampler (preprocessor_config and resampler_config) by @istupakov in #48
• Added concurrent processing in preprocessor (set max_concurrent_workers in preprocessor_config) by @istupakov in #47, #48
• Fixed VAD for 8 kHz models (t-tech/t-one) by @istupakov in #46
• CTC decoding optimization by @istupakov in #47

Full Changelog: v0.9.0...v0.9.1

onnx-asr v0.9.0

Highlights

• New supported model nemo-canary-1b-v2 for Nvidia Canary 1B v2 - one of the best multilingual models according to open_asr_leaderboard so far!
• New supported model t-tech/t-one for T-Tech T-one - Russian model, specialized for the telephony domain.

What's Changed

• Add Nemo Canary models support by @istupakov in #42
• Add T-one model by @istupakov in #44
• Improve resampling by @istupakov in #43
• Update dependencies, linter and readme by @istupakov in #41, #45

Full Changelog: v0.8.0...v0.9.0

onnx-asr v0.8.0

Highlights

• Added support for new GigaAM v3 models, including E2E versions with punctuation and text normalization!

What's Changed

• Add GigaAM v3 models by @istupakov in #34
• Update GigaAM and Nemo preprocessors by @istupakov in #36
• Update benchmarks by @istupakov in #38
• Add GigaAM v3 E2E models by @istupakov in #39

Full Changelog: v0.7.0...v0.8.0

onnx-asr v0.7.0

Highlights

• New supported model nemo-parakeet-tdt-0.6b-v3 for Nvidia Parakeet TDT 0.6B V3 - new multilingual Parakeet model from Nvidia!
• Performance optimization for GigaAM v2 RNN-T and WhisperHf models.
• Added performance benchmarks in Readme (on Arm, Intel, and Nvidia T4).

What's Changed

• Update deps and onnxscript by @istupakov in #15, #19, #20
• Fix Tdt decoding by @istupakov in #14
• Performance optimization by @istupakov in #17
• Update benchmarks by @istupakov in #16
• Add parakeet-tdt-0.6b-v3 model by @istupakov in #25

Full Changelog: v0.6.1...v0.7.0

onnx-asr v0.6.1

What's Changed

• New supported sample rates (24 and 32 kHz)
• Set cpu_preprocessing to true by default
• Preprocessors optimization by @istupakov in #10
• Update Readme by @istupakov in #12

Full Changelog: v0.6.0...v0.6.1

onnx-asr v0.6.0

Highlights

• New supported model nemo-parakeet-tdt-0.6b-v2 for Nvidia Parakeet TDT 0.6B V2 (en) - the best English model according to open_asr_leaderboard at the moment
• load_model now supports loading from any Hugging Face repository with a compatible config.json
• All models now have a quantized version (quantization="uint8" for onnx-community models and quantization="int8" for others)

What's Changed

• Improved typing, updated CLI by @istupakov in #5
• Optimized preprocessors on GPU and added cpu_preprocessing option to load_model by @istupakov in #6
• Added support for Nemo TDT models by @istupakov in #4
• Added config.json support by @istupakov in #8
• Improved tests and typings by @istupakov in #9
• [Breaking Changes] Renamed model type whisper-hf to whisper (for compatibility with config.json from onnx-community)

Full Changelog: v0.5.0...v0.6.0