This is an automated email from the git hooks/post-receive script.
guix_mirror_bot pushed a commit to branch master
in repository guix.
The following commit(s) were added to refs/heads/master by this push:
new 53ba4e4646 gnu: Add python-onnx-asr.
53ba4e4646 is described below
commit 53ba4e46467782fca0e0658ee07751f671025b5b
Author: Danny Milosavljevic <[email protected]>
AuthorDate: Sat Mar 7 04:12:00 2026 +0100
gnu: Add python-onnx-asr.
* gnu/packages/patches/python-onnx-asr-0.10.2-bundled-parakeet-model.patch:
New file.
* gnu/local.mk (dist_patch_DATA): Add reference to it.
* gnu/packages/speech.scm (%parakeet-hf-base): New variable.
(parakeet-tdt-config): New variable.
(parakeet-tdt-vocab): New variable.
(parakeet-tdt-encoder-int8): New variable.
(parakeet-tdt-decoder-joint-int8): New variable.
(python-onnx-asr): New variable.
Change-Id: Id103dfe51478b68d8332308fa30e12e5f16d1d80
---
gnu/local.mk | 1 +
...on-onnx-asr-0.10.2-bundled-parakeet-model.patch | 21 ++++
gnu/packages/speech.scm | 129 +++++++++++++++++++++
3 files changed, 151 insertions(+)
diff --git a/gnu/local.mk b/gnu/local.mk
index beff7a9300..6267e56038 100644
--- a/gnu/local.mk
+++ b/gnu/local.mk
@@ -2101,6 +2101,7 @@ dist_patch_DATA =
\
%D%/packages/patches/python-msal-requests.patch \
%D%/packages/patches/python-norns-nose.patch \
%D%/packages/patches/python-numpy-gcc-14.patch \
+ %D%/packages/patches/python-onnx-asr-0.10.2-bundled-parakeet-model.patch
\
%D%/packages/patches/python-random2-getrandbits-test.patch \
%D%/packages/patches/python-pillow-use-zlib-1.3.patch \
%D%/packages/patches/python-pydocstyle-add-support-for-pep701.patch \
diff --git
a/gnu/packages/patches/python-onnx-asr-0.10.2-bundled-parakeet-model.patch
b/gnu/packages/patches/python-onnx-asr-0.10.2-bundled-parakeet-model.patch
new file mode 100644
index 0000000000..b6516e3c49
--- /dev/null
+++ b/gnu/packages/patches/python-onnx-asr-0.10.2-bundled-parakeet-model.patch
@@ -0,0 +1,21 @@
+From: Danny Milosavljevic <[email protected]>
+Date: 2026-02-13
+Subject: Use bundled Parakeet TDT V3 int8 model instead of downloading from
+ HuggingFace.
+
+The @PARAKEET_MODEL_DIR@ placeholder is replaced with the actual store path
+during the build.
+
+--- a/src/onnx_asr/loader.py
++++ b/src/onnx_asr/loader.py
+@@ -308,6 +308,10 @@
+ case "nemo-parakeet-tdt-0.6b-v3":
+ model_type = NemoConformerTdt
+ default_repo_id = "istupakov/parakeet-tdt-0.6b-v3-onnx"
++ if path is None:
++ path = "@PARAKEET_MODEL_DIR@"
++ if quantization is None:
++ quantization = "int8"
+ case "nemo-conformer-aed":
+ model_type = NemoConformerAED
+ case "nemo-canary-1b-v2":
diff --git a/gnu/packages/speech.scm b/gnu/packages/speech.scm
index e771a29520..38858a6e68 100644
--- a/gnu/packages/speech.scm
+++ b/gnu/packages/speech.scm
@@ -824,3 +824,132 @@ intermediate representation.")
(description "This package enables developers to author ONNX models
using a Python-based domain-specific language.")
(license license:expat)))
+
+;;; Parakeet TDT V3 ONNX model weights from istupakov/parakeet-tdt-0.6b-v3-onnx
+;;; (HuggingFace, revision abd2878d52a678ce380088ef9d9b1d9664404565).
+;;; License: CC-BY-4.0 (NVIDIA).
+;;; Int8 quantized variant (~670 MB total).
+
+(define %parakeet-hf-base
+
"https://huggingface.co/istupakov/parakeet-tdt-0.6b-v3-onnx/resolve/abd2878d52a678ce380088ef9d9b1d9664404565";)
+
+(define parakeet-tdt-config
+ (origin
+ (method url-fetch)
+ (uri (string-append %parakeet-hf-base "/config.json"))
+ (file-name "config.json")
+ (sha256
+ (base32 "0rn4i8ad5h1vga6yq04qpy6qmc30rpvd9bqhqbrcm64pdg3h6sb6"))))
+
+(define parakeet-tdt-vocab
+ (origin
+ (method url-fetch)
+ (uri (string-append %parakeet-hf-base "/vocab.txt"))
+ (file-name "vocab.txt")
+ (sha256
+ (base32 "0pf3wcvps76wq7iadw37lk7xcjs7gpmlbxficg2nmg54krkl91fm"))))
+
+(define parakeet-tdt-encoder-int8
+ (origin
+ (method url-fetch)
+ (uri (string-append %parakeet-hf-base "/encoder-model.int8.onnx"))
+ (file-name "encoder-model.int8.onnx")
+ (sha256
+ (base32 "02gzb82y86vl7jr69bn7qyfbifpd4nbi9ivpnabn020vgvxd4fb1"))))
+
+(define parakeet-tdt-decoder-joint-int8
+ (origin
+ (method url-fetch)
+ (uri (string-append %parakeet-hf-base "/decoder_joint-model.int8.onnx"))
+ (file-name "decoder_joint-model.int8.onnx")
+ (sha256
+ (base32 "0w3scrvqj74xv6h2f8c1k2q9234nwf1yvj7dv9sh78yiwcz4i9zf"))))
+
+(define-public python-onnx-asr
+ (package
+ (name "python-onnx-asr")
+ (version "0.10.2")
+ (source
+ (origin
+ (method url-fetch)
+ (uri (pypi-uri "onnx_asr" version))
+ (sha256
+ (base32 "0d5vmkavcqjf7b2aa0nc118b2pf34mc7yzjkaw92rl42rwwijf3h"))
+ (patches
+ (search-patches
"python-onnx-asr-0.10.2-bundled-parakeet-model.patch"))))
+ (build-system pyproject-build-system)
+ (arguments
+ (list
+ #:test-flags
+ #~(list ;; These tests try to download models from HuggingFace Hub.
+ "--ignore=tests/onnx_asr/test_recognize.py"
+ "--ignore=tests/onnx_asr/test_cli.py"
+ "--ignore=tests/onnx_asr/test_load_model_errors.py"
+ ;;; These tests would compare preprocessor output against
+ ;;; reference implementations that are not in Guix:
+ ;;; - kaldi_native_fbank (C++ lib, not packaged)
+ ;;; - nemo (NVIDIA NeMo framework, not packaged)
+ ;;; - openai-whisper (Python package, not packaged;
+ ;;; whisper-cpp exists but is C++ only, no Python module)
+ "--ignore=tests/preprocessors/test_kaldi.py"
+ "--ignore=tests/preprocessors/test_nemo.py"
+ "--ignore=tests/preprocessors/test_whisper_preprocessor.py")
+ #:phases
+ #~(modify-phases %standard-phases
+ (add-after 'install 'install-parakeet-model
+ (lambda* (#:key inputs outputs #:allow-other-keys)
+ (let* ((out (assoc-ref outputs "out"))
+ (site (string-append out "/lib/python"
+ #$(version-major+minor
+ (package-version python))
+ "/site-packages/onnx_asr"))
+ (model-dir (string-append site
+ "/models-data"
+ "/parakeet-tdt-0.6b-v3")))
+ (mkdir-p model-dir)
+ (symlink (assoc-ref inputs "config.json")
+ (string-append model-dir "/config.json"))
+ (symlink (assoc-ref inputs "vocab.txt")
+ (string-append model-dir "/vocab.txt"))
+ (symlink (assoc-ref inputs "encoder-model.int8.onnx")
+ (string-append model-dir
+ "/encoder-model.int8.onnx"))
+ (symlink (assoc-ref inputs
+ "decoder_joint-model.int8.onnx")
+ (string-append model-dir
+ "/decoder_joint-model.int8.onnx")))))
+ (add-after 'install-parakeet-model 'patch-model-paths
+ (lambda* (#:key outputs #:allow-other-keys)
+ (let* ((out (assoc-ref outputs "out"))
+ (site (string-append out "/lib/python"
+ #$(version-major+minor
+ (package-version python))
+ "/site-packages/onnx_asr"))
+ (model-dir (string-append site
+ "/models-data"
+ "/parakeet-tdt-0.6b-v3")))
+ (substitute* (string-append site "/loader.py")
+ (("@PARAKEET_MODEL_DIR@") model-dir))))))))
+ (propagated-inputs
+ (list python-numpy
+ python-huggingface-hub
+ (list onnxruntime "python")))
+ (native-inputs
+ (list nss-certs-for-test
+ onnx
+ parakeet-tdt-config
+ parakeet-tdt-vocab
+ parakeet-tdt-encoder-int8
+ parakeet-tdt-decoder-joint-int8
+ python-hatchling
+ python-onnxscript
+ python-pytorch
+ python-pytest
+ python-torchaudio))
+ (home-page "https://github.com/istupakov/onnx-asr";)
+ (synopsis "Speech recognition using ONNX models")
+ (description
+ "ONNX ASR is a Python library for automatic speech recognition using
+ONNX Runtime. It supports models including Whisper and NeMo Parakeet.
+Includes bundled Parakeet TDT V3 model weights (int8, CC-BY-4.0, NVIDIA).")
+ (license license:expat)))
