Add validated Parakeet TDT v3 ONNX INT8 deployment format

onnx/README.md

@@ -0,0 +1,93 @@
+# Orukeet for sherpa-onnx
+
+The released Orukeet r3 checkpoint in the standard **Parakeet TDT v3 ONNX layout**.
+Use it with sherpa-onnx's existing offline-transducer loader. The fitted Gabor
+filters are ordinary convolution weights; this export requires no
+Orukeet-specific operator or frontend.
+
+[Download the INT8 package](https://huggingface.co/oruk/orukeet/resolve/main/onnx/sherpa-onnx-orukeet-v0.1.0-int8.tar.bz2)
+· [Model card](https://huggingface.co/oruk/orukeet)
+· [Export code](https://github.com/Oruk-AI/orukeet/tree/main/export/onnx)
+· [Validation receipts](https://github.com/Oruk-AI/orukeet/tree/main/evidence/onnx-r3-20260910)
+
+The download is **486,664,389 bytes** and extracts to **670,718,459 bytes**.
+Archive SHA-256:
+
+```text
+c4ad85bbfb0835167c097dedcec0bb3f50cc468edfe7e690442e2221427b95bf
+```
+
+```text
+sherpa-onnx-orukeet-v0.1.0-int8/
+  encoder.int8.onnx
+  decoder.int8.onnx
+  joiner.int8.onnx
+  tokens.txt
+  bpe.vocab
+  LICENSE-WEIGHTS
+  NOTICE.md
+```
+
+## Transcribe
+
+Extract the archive, then install the runtime:
+
+```sh
+tar -xjf sherpa-onnx-orukeet-v0.1.0-int8.tar.bz2
+python -m pip install sherpa-onnx==1.13.4 soundfile numpy
+```
+
+```python
+from pathlib import Path
+
+import sherpa_onnx
+import soundfile as sf
+
+model = Path("sherpa-onnx-orukeet-v0.1.0-int8")
+recognizer = sherpa_onnx.OfflineRecognizer.from_transducer(
+    encoder=str(model / "encoder.int8.onnx"),
+    decoder=str(model / "decoder.int8.onnx"),
+    joiner=str(model / "joiner.int8.onnx"),
+    tokens=str(model / "tokens.txt"),
+    model_type="nemo_transducer",
+    sample_rate=16000,
+    feature_dim=128,
+    decoding_method="greedy_search",
+    provider="cpu",
+    num_threads=4,
+)
+
+samples, sample_rate = sf.read("audio.wav", dtype="float32", always_2d=True)
+stream = recognizer.create_stream()
+stream.accept_waveform(sample_rate, samples.mean(axis=1))
+recognizer.decode_stream(stream)
+print(stream.result.text)
+```
+
+Applications can retain their existing VAD and chunking. For long recordings,
+use bounded speech segments. The model retains full-context Parakeet attention.
+
+## Export details
+
+The graphs use opset 17, unsigned INT8 encoder quantization, and signed INT8
+decoder/joiner quantization. The decoder uses ONNX Runtime's
+`DynamicQuantizeLSTM`, as the stock Parakeet export does. The original
+128-bin features, 8× subsampling, 8,192-token vocabulary plus blank, and TDT
+durations `[0, 1, 2, 3, 4]` are retained.
+
+sherpa-onnx 1.13.4 identifies TDT through `tdt` in the encoder's `url` metadata.
+This export sets that field to
+`https://huggingface.co/oruk/orukeet#parakeet-tdt-v3`; keep it when repackaging.
+
+Source checkpoint SHA-256:
+
+```text
+031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56
+```
+
+[Reproduce the export](https://github.com/Oruk-AI/orukeet/blob/main/export/onnx/README.md)
+· [Per-file hashes](https://github.com/Oruk-AI/orukeet/blob/main/evidence/onnx-r3-20260910/package-manifest.json)
+
+The ONNX weights inherit **CC BY-SA 4.0** from the released Orukeet checkpoint,
+including attribution to NVIDIA Parakeet TDT 0.6B v3. The complete license and
+attribution are included in the archive.

onnx/manifest.json

@@ -0,0 +1,46 @@
+{
+  "source_sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56",
+  "archive": "sherpa-onnx-orukeet-v0.1.0-int8.tar.bz2",
+  "archive_bytes": 486664389,
+  "archive_sha256": "c4ad85bbfb0835167c097dedcec0bb3f50cc468edfe7e690442e2221427b95bf",
+  "extract_dir": "sherpa-onnx-orukeet-v0.1.0-int8",
+  "files": [
+    {
+      "path": "encoder.int8.onnx",
+      "bytes": 652282402,
+      "sha256": "d10711f1b8f3a516e2d7a93adb219caf8aba2b55295305db92f3e80e78c1499a"
+    },
+    {
+      "path": "decoder.int8.onnx",
+      "bytes": 11845332,
+      "sha256": "c185c2afb4c77c94bb1314807ecb3dc1623057a3dc540b83e10301af9bf4cfca"
+    },
+    {
+      "path": "joiner.int8.onnx",
+      "bytes": 6355335,
+      "sha256": "1a7e90abf7172d926dd7e2edac2a5d5035c24dfb641a15e131d57b6a5f63cdd3"
+    },
+    {
+      "path": "tokens.txt",
+      "bytes": 93939,
+      "sha256": "d58544679ea4bc6ac563d1f545eb7d474bd6cfa467f0a6e2c1dc1c7d37e3c35d"
+    },
+    {
+      "path": "bpe.vocab",
+      "bytes": 117408,
+      "sha256": "41d5e71b3591642eff088151efd7acd4e750124cc0054c8ba9fa3245187a4804"
+    },
+    {
+      "path": "LICENSE-WEIGHTS",
+      "bytes": 20137,
+      "sha256": "23ee78c8bae49cf08ea2f0c84945c66b987ebe4520881fb51b3dad4fb43d07c2"
+    },
+    {
+      "path": "NOTICE.md",
+      "bytes": 3906,
+      "sha256": "c1c90ecec7c9ff8056aada8979bf4f3f4dba901c54297c14a328815160e20c6e"
+    }
+  ],
+  "export_receipt_sha256": "9fc487cdfa076b4d210a526c68ff24ea9ff7ff9da78e058b21f7d610f95ccc6e",
+  "packager_sha256": "8acbfe6475a5ea3144311127a04b1cc319cac6a875835bce7ac6d50c19de9f9d"
+}

onnx/sherpa-onnx-orukeet-v0.1.0-int8.tar.bz2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4ad85bbfb0835167c097dedcec0bb3f50cc468edfe7e690442e2221427b95bf
+size 486664389