shruti-onnx 0.2.0__tar.gz

shruti_onnx-0.2.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

shruti_onnx-0.2.0/PKG-INFO

@@ -0,0 +1,46 @@
+Metadata-Version: 2.4
+Name: shruti-onnx
+Version: 0.2.0
+Summary: Shruti: offline speech-to-text, one language and English in one small model, in ONNX Runtime (32 MB).
+Author: Udgam Labs
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://huggingface.co/udgamlabs/Shruti-ONNX
+Project-URL: Source, https://github.com/udgamlabs/shruti
+Keywords: speech-recognition,speech-to-text,onnx,offline,on-device
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy
+Requires-Dist: onnxruntime>=1.17
+Requires-Dist: huggingface_hub
+Provides-Extra: audio
+Requires-Dist: soundfile; extra == "audio"
+Requires-Dist: soxr; extra == "audio"
+Dynamic: license-file
+
+# shruti-onnx
+
+Shruti: offline speech-to-text, one small model per language. Each model covers its language, English, and the two mixed. Models and full results: [udgamlabs/Shruti-ONNX](https://huggingface.co/udgamlabs/Shruti-ONNX).
+
+Languages: Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu, Urdu. Use the word or the code.
+
+```bash
+pip install shruti-onnx
+```
+
+```python
+from shruti_onnx import Shruti
+
+asr = Shruti.load("urdu")                    # or "ur"; variant="int8" for low memory on ARM
+print(asr.transcribe("clip.wav")["text"])    # language detected; language="en" to set English
+```
+
+Command line: `shruti-onnx clip.wav --lang urdu`. `loops="cut"` removes a repeat and what follows it.
+
+## Where it falls short
+
+Nine of these languages (Dogri, Hindi, Kannada, Konkani, Odia, Punjabi, Sindhi, Tamil, Telugu) failed our written comparison against the original engine, and none met every rule. The model card names each miss and its size. On Linux x86, use the default 4-bit build: the int8 and sherpa-onnx builds were less accurate there on the machine we tested. A model can get stuck repeating a phrase; the text stays as the model wrote it unless you ask to cut it. The 4-bit build needs about 1 GB of memory once loaded in ONNX Runtime.
+
+Apache-2.0. Built on [Cactus Whistle](https://huggingface.co/Cactus-Compute/whistle).
+
+**About Udgam Labs.** Udgam Labs is an independent AI research lab. [udgamlabs.com](https://udgamlabs.com)

shruti_onnx-0.2.0/README.md

@@ -0,0 +1,26 @@
+# shruti-onnx
+
+Shruti: offline speech-to-text, one small model per language. Each model covers its language, English, and the two mixed. Models and full results: [udgamlabs/Shruti-ONNX](https://huggingface.co/udgamlabs/Shruti-ONNX).
+
+Languages: Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu, Urdu. Use the word or the code.
+
+```bash
+pip install shruti-onnx
+```
+
+```python
+from shruti_onnx import Shruti
+
+asr = Shruti.load("urdu")                    # or "ur"; variant="int8" for low memory on ARM
+print(asr.transcribe("clip.wav")["text"])    # language detected; language="en" to set English
+```
+
+Command line: `shruti-onnx clip.wav --lang urdu`. `loops="cut"` removes a repeat and what follows it.
+
+## Where it falls short
+
+Nine of these languages (Dogri, Hindi, Kannada, Konkani, Odia, Punjabi, Sindhi, Tamil, Telugu) failed our written comparison against the original engine, and none met every rule. The model card names each miss and its size. On Linux x86, use the default 4-bit build: the int8 and sherpa-onnx builds were less accurate there on the machine we tested. A model can get stuck repeating a phrase; the text stays as the model wrote it unless you ask to cut it. The 4-bit build needs about 1 GB of memory once loaded in ONNX Runtime.
+
+Apache-2.0. Built on [Cactus Whistle](https://huggingface.co/Cactus-Compute/whistle).
+
+**About Udgam Labs.** Udgam Labs is an independent AI research lab. [udgamlabs.com](https://udgamlabs.com)

shruti_onnx-0.2.0/pyproject.toml

@@ -0,0 +1,24 @@
+[build-system]
+requires = ["setuptools>=77"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "shruti-onnx"
+version = "0.2.0"
+description = "Shruti: offline speech-to-text, one language and English in one small model, in ONNX Runtime (32 MB)."
+requires-python = ">=3.9"
+license = "Apache-2.0"
+readme = "README.md"
+authors = [{ name = "Udgam Labs" }]
+keywords = ["speech-recognition", "speech-to-text", "onnx", "offline", "on-device"]
+dependencies = ["numpy", "onnxruntime>=1.17", "huggingface_hub"]
+
+[project.optional-dependencies]
+audio = ["soundfile", "soxr"]
+
+[project.urls]
+Homepage = "https://huggingface.co/udgamlabs/Shruti-ONNX"
+Source = "https://github.com/udgamlabs/shruti"
+
+[project.scripts]
+shruti-onnx = "shruti_onnx:main"

shruti_onnx-0.2.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

shruti_onnx-0.2.0/shruti_onnx/__init__.py

@@ -0,0 +1,297 @@
+"""Shruti: offline speech-to-text, one language and English in one small model, in ONNX Runtime.
+
+    from shruti_onnx import Shruti
+    asr = Shruti.load("tamil")                # a plain word or a code ("ta"); see LANGUAGES
+    print(asr.transcribe("clip.wav")["text"])
+
+    Shruti.load("ta", variant="int8")         # low memory; the default is the 4-bit build, the smallest download
+    asr.transcribe(samples_16k, language="en")   # None (the default) detects: the model's language or English
+
+Decoding is 5-beam search with length-normalised scores, which reproduces the Cactus engine's transcripts. The
+search is the same as the research runner udgam-voice/asr/onnx/shruti_onnx.py (tested equal in tests/).
+"""
+import json
+import os
+import wave
+
+import numpy as np
+import onnxruntime as ort
+
+__version__ = "0.2.0"
+DEFAULT_REPO = "udgamlabs/Shruti-ONNX"
+BASE_TAGS = ("en", "de", "fr", "es", "it", "nl", "pl")  # Cactus Whistle's tags; a Shruti model adds its own language
+VARIANTS = {"4bit": "onnx", "int8": "onnx-int8"}
+_ALL_LANGUAGES = {"as": "Assamese", "bn": "Bengali", "brx": "Bodo", "doi": "Dogri", "gu": "Gujarati", "hi": "Hindi", "kn": "Kannada", "ks": "Kashmiri",
+             "kok": "Konkani", "mai": "Maithili", "ml": "Malayalam", "mni": "Manipuri", "mr": "Marathi", "ne": "Nepali", "or": "Odia", "pa": "Punjabi",
+             "sa": "Sanskrit", "sat": "Santali", "sd": "Sindhi", "ta": "Tamil", "te": "Telugu", "ur": "Urdu"}
+RELEASED = ("as", "bn", "brx", "doi", "gu", "hi", "kn", "kok", "ks", "mai", "ml", "mni", "mr", "ne", "or", "pa", "sa", "sat", "sd", "ta", "te", "ur")  # set by set_release.py
+LANGUAGES = {c: _ALL_LANGUAGES[c] for c in RELEASED}
+_ALIASES = {"bangla": "bn", "oriya": "or", "meitei": "mni", "meetei": "mni", "panjabi": "pa"}
+LOOPS_DEFAULT = "off"
+RATE = 16000
+
+
+def language_code(name):
+    """"urdu", "Urdu" or "ur" -> "ur". Raises ValueError listing the languages."""
+    key = str(name).strip().lower()
+    if key in _ALL_LANGUAGES:
+        return key
+    for code, word in _ALL_LANGUAGES.items():
+        if key == word.lower():
+            return code
+    if key in _ALIASES:
+        return _ALIASES[key]
+    raise ValueError(f"unknown language {name!r}; choose one of: " + ", ".join(f"{w} ({c})" for c, w in sorted(LANGUAGES.items(), key=lambda x: x[1])))
+
+
+def read_audio(path):
+    """16 kHz mono float32 samples from a file: soundfile when installed, else 16/24/32-bit PCM WAV via the standard
+    library. Other sample rates are resampled (soxr when installed, else linear interpolation)."""
+    try:
+        import soundfile as sf
+        x, rate = sf.read(path, dtype="float32", always_2d=True)
+        x = x.mean(1)
+    except ImportError:
+        with wave.open(path, "rb") as w:
+            ch, width, rate, n = w.getnchannels(), w.getsampwidth(), w.getframerate(), w.getnframes()
+            raw = w.readframes(n)
+        if width == 3:
+            b = np.frombuffer(raw, np.uint8).reshape(-1, 3)
+            x = (b[:, 0].astype(np.int32) | (b[:, 1].astype(np.int32) << 8) | (b[:, 2].astype(np.int8).astype(np.int32) << 16))
+            x = x / 8388608.0
+        else:
+            x = np.frombuffer(raw, {1: np.uint8, 2: np.int16, 4: np.int32}[width]).astype(np.float64)
+            x = (x - 128) / 128.0 if width == 1 else x / float(2 ** (8 * width - 1))
+        x = x.reshape(-1, ch).mean(1).astype(np.float32)
+    return resample(x, rate)
+
+
+def resample(x, rate):
+    if rate == RATE:
+        return np.asarray(x, np.float32)
+    try:
+        import soxr
+        return soxr.resample(x, rate, RATE).astype(np.float32)
+    except ImportError:
+        n = int(round(len(x) * RATE / rate))
+        return np.interp(np.arange(n) * rate / RATE, np.arange(len(x)), x).astype(np.float32)
+
+
+class Shruti:
+    """Runs the three graphs: frontend (audio -> features), encoder (features -> audio memory), decoder (one token
+    at a time, with its state carried between calls)."""
+
+    def __init__(self, folder, threads=2):
+        so = ort.SessionOptions()
+        so.intra_op_num_threads = threads
+        so.inter_op_num_threads = 1
+        load = lambda f: ort.InferenceSession(os.path.join(folder, f), so, providers=["CPUExecutionProvider"])  # noqa
+        self.frontend, self.encoder, self.decoder = load("frontend.onnx"), load("encoder.onnx"), load("decoder.onnx")
+        meta = json.load(open(os.path.join(folder, "shruti.json"), encoding="utf-8"))
+        self.vocab, self.cfg = meta["vocab"], meta["config"]
+        self.text, self.bos, self.eos = self.cfg["text"], self.cfg["bos"], self.cfg["eos"]
+        self.tags = {lang: self.text + i for i, lang in enumerate(self.cfg["languages"])}
+        own = [x for x in self.cfg["languages"] if x not in BASE_TAGS]
+        self.languages = (own[0], "en")  # the model's own language and English: what it was trained on
+        self.state_names = [i.name for i in self.decoder.get_inputs()][2:]
+        shapes = {i.name: i.shape for i in self.decoder.get_inputs()}
+        self.state_shapes = {n: [d if isinstance(d, int) else 0 for d in shapes[n]] for n in self.state_names}
+
+    @classmethod
+    def load(cls, language, variant="4bit", model=DEFAULT_REPO, threads=2):
+        """language: a plain word or a code ("urdu" or "ur"); see LANGUAGES.
+        variant: "4bit" (32 MB, the smallest download) or "int8" (53 MB, about 120 MB of memory instead of about 1 GB).
+        model: a Hugging Face repo id, or a local folder holding <code>/onnx/ and <code>/onnx-int8/."""
+        code = language_code(language)
+        if code not in LANGUAGES and not os.path.isdir(model):
+            raise ValueError(f"{_ALL_LANGUAGES[code]} is not in this release; choose one of: " + ", ".join(f"{w} ({c})" for c, w in sorted(LANGUAGES.items(), key=lambda x: x[1])))
+        if variant not in VARIANTS:
+            raise ValueError(f"variant must be one of {list(VARIANTS)}, not {variant!r}")
+        rel = f"{code}/{VARIANTS[variant]}"
+        if os.path.isdir(model):
+            path = model
+        else:
+            from huggingface_hub import snapshot_download
+            path = snapshot_download(model, allow_patterns=[f"{rel}/*"])
+        if not os.path.isdir(os.path.join(path, rel)):
+            raise FileNotFoundError(f"no {rel} folder in {model}: this language is not in this release; see its table of languages")
+        asr = cls(os.path.join(path, rel), threads)
+        if asr.languages[0] != code:
+            raise ValueError(f"{rel} holds a {asr.languages[0]!r} model, not {code!r}")
+        return asr
+
+    def detok(self, ids):
+        """Pieces to text (sentencepiece rules: a leading ▁ is a space, <0xNN> pieces are raw UTF-8 bytes)."""
+        out, raw = [], bytearray()
+        for i in ids:
+            p = self.vocab[i]
+            if len(p) == 6 and p.startswith("<0x") and p.endswith(">"):
+                raw.append(int(p[3:5], 16))
+                continue
+            if raw:
+                out.append(raw.decode("utf-8", "ignore"))
+                raw.clear()
+            if not (p.startswith("<") and p.endswith(">")):
+                out.append(p)
+        if raw:
+            out.append(raw.decode("utf-8", "ignore"))
+        return "".join(out).replace("▁", " ").strip()
+
+    def init_state(self, B):
+        st = {}
+        for n in self.state_names:
+            shp = list(self.state_shapes[n])
+            if n == "pos":
+                st[n] = np.zeros(1, np.int64)
+                continue
+            shp[0] = B
+            st[n] = np.zeros(shp, np.int64 if n == "hist" else np.float32)
+        return st
+
+    def step(self, tokens, cross, st):
+        r = self.decoder.run(None, {"tokens": tokens, "cross": cross, **st})
+        return r[0][:, 0], dict(zip(self.state_names, r[1:]))
+
+    def encode(self, wav):
+        feats = self.frontend.run(None, {"audio": wav[None].astype(np.float32)})[0]
+        n = np.array([len(wav) // self.cfg["hop"]], np.int64)
+        return self.encoder.run(None, {"features": feats, "frames": n})[0]
+
+    @staticmethod
+    def logsoftmax(x):
+        x = x - x.max(-1, keepdims=True)
+        return x - np.log(np.exp(x).sum(-1, keepdims=True))
+
+    def transcribe(self, audio, language=None, beams=5, numbers="words", window_s=30.0, length_norm=True,
+                   tok_per_s=12.0, loops=LOOPS_DEFAULT):
+        """audio: a file path, or 16 kHz mono float32 samples in [-1, 1].
+        language: None detects between the model's own language and English; or name either one ("tamil", "ta", "english", "en").
+        beams: 5 (the engine's search) or 1 (greedy, faster).
+        numbers: "words" (as spoken: "ढाई सौ रुपये") or "digits" ("₹250"; Hindi model only, needs the
+        udgam_textnorm normaliser, not on PyPI yet).
+        loops: "off" keeps the model's output as it is; "cut" keeps the first copy of a repeated pattern and drops
+        everything after it.
+        Up to 30 s is one pass; longer audio is cut into equal windows of at most 30 s.
+        -> {"text", "language", "tokens"}"""
+        if language is not None:
+            language = "en" if str(language).strip().lower() in ("en", "english") else language_code(language)
+            if language not in self.languages:
+                raise ValueError(f"this is the {LANGUAGES[self.languages[0]]} model: language must be {self.languages[0]!r}, 'en' or None to detect, not {language!r}")
+        if numbers == "digits" and self.languages[0] != "hi":
+            raise ValueError('numbers="digits" is tested for the Hindi model only')
+        if loops not in ("off", "cut"):
+            raise ValueError('loops must be "off" or "cut"')
+        wav = read_audio(audio) if isinstance(audio, (str, os.PathLike)) else np.asarray(audio, np.float32)
+        if len(wav) == 0 or np.abs(wav).max() < 1e-4:  # digital silence -> empty, as the Cactus engine returns
+            return {"text": "", "language": language, "tokens": []}
+        n = max(1, int(np.ceil(len(wav) / (window_s * RATE))))
+        parts = [self._transcribe(w, language, beams, tok_per_s, length_norm, loops) for w in np.array_split(wav, n)]
+        text = " ".join(p["text"] for p in parts if p["text"])
+        lang = parts[0]["language"]
+        if numbers == "digits":
+            try:
+                from udgam_textnorm import denormalize
+            except ImportError as e:
+                raise ImportError('numbers="digits" needs the udgam_textnorm number normaliser, which is not on PyPI yet') from e
+            text = denormalize(text, lang, style="digits")
+        elif numbers != "words":
+            raise ValueError('numbers must be "words" or "digits"')
+        return {"text": text, "language": lang, "tokens": [t for p in parts for t in p["tokens"]]}
+
+    @staticmethod
+    def cut_loops(ids, max_period=8, repeats=3):
+        """Cut a token loop (a pattern of up to 8 tokens repeated 3+ times back to back): keep its first copy."""
+        n = len(ids)
+        for start in range(n):
+            for k in range(1, max_period + 1):
+                if start + k * repeats > n:
+                    break
+                pat = ids[start:start + k]
+                if all(ids[start + r * k:start + (r + 1) * k] == pat for r in range(1, repeats)):
+                    return ids[:start + k]
+        return ids
+
+    def _transcribe(self, wav, language, beams, tok_per_s, length_norm, loops):
+        tidy = self.cut_loops if loops == "cut" else (lambda ids: ids)
+        cross = self.encode(wav)
+        cap = int(min(self.cfg["max_tokens"] - 2, len(wav) / RATE * tok_per_s + 8))
+        st = self.init_state(1)
+        if language is None:
+            lg, st = self.step(np.array([[self.bos]], np.int64), cross, st)
+            language = max(self.languages, key=lambda lang: lg[0, self.tags[lang]])
+            lg, st = self.step(np.array([[self.tags[language]]], np.int64), cross, st)
+        else:
+            lg, st = self.step(np.array([[self.bos, self.tags[language]]], np.int64), cross, st)
+        lg[:, self.text:] = -np.inf
+        if beams <= 1:
+            ids = []
+            for _ in range(cap):
+                k = int(lg[0].argmax())
+                if k == self.eos:
+                    break
+                ids.append(k)
+                lg, st = self.step(np.array([[k]], np.int64), cross, st)
+                lg[:, self.text:] = -np.inf
+            ids = tidy(ids)
+            return {"text": self.detok(ids), "language": language, "tokens": ids}
+        norm = (lambda sc, s: sc / (len(s) + 1)) if length_norm else (lambda sc, s: sc)
+        lp = self.logsoftmax(lg[0])
+        first = np.argsort(-lp)[:beams]
+        seqs = [[int(k)] for k in first]
+        scores = lp[first].astype(np.float64)
+        finished = [(scores[i], s[:-1]) for i, s in enumerate(seqs) if s[-1] == self.eos]
+        live = [i for i, s in enumerate(seqs) if s[-1] != self.eos]
+        seqs, scores = [seqs[i] for i in live], scores[live]
+        st = {n: (v if n == "pos" else np.repeat(v, len(seqs), 0)) for n, v in st.items()}
+        for _ in range(cap - 1):
+            if not seqs:
+                break
+            lg, st = self.step(np.array([[s[-1]] for s in seqs], np.int64), cross, st)
+            lg[:, self.text:] = -np.inf
+            cand = scores[:, None] + self.logsoftmax(lg)
+            new_seqs, new_scores, src = [], [], []
+            for flat in np.argsort(-cand, axis=None)[: 2 * beams]:
+                b, k = divmod(int(flat), cand.shape[1])
+                if k == self.eos:
+                    finished.append((cand[b, k], seqs[b]))
+                else:
+                    new_seqs.append(seqs[b] + [k])
+                    new_scores.append(cand[b, k])
+                    src.append(b)
+                if len(new_seqs) == beams:
+                    break
+            best_done = max((norm(sc, s) for sc, s in finished), default=-np.inf)
+            if len(finished) >= beams and (not new_scores or
+                                           max(norm(sc, s) for sc, s in zip(new_scores, new_seqs)) < best_done):
+                break
+            seqs, scores = new_seqs, np.array(new_scores)
+            st = {n: (v if n == "pos" else v[src]) for n, v in st.items()}
+        if not finished:
+            finished = list(zip(scores, seqs))
+        best = tidy(max(finished, key=lambda f: norm(*f))[1])
+        return {"text": self.detok(best), "language": language, "tokens": best}
+
+
+def main(argv=None):
+    import argparse
+    p = argparse.ArgumentParser(prog="shruti-onnx", description="Offline speech-to-text: one language and English in one small model.")
+    p.add_argument("audio", nargs="+", help="audio files (WAV; other formats with `pip install soundfile`)")
+    p.add_argument("--lang", required=True, help="which model: a plain word or a code, e.g. tamil or ta")
+    p.add_argument("--int8", action="store_true", help="the low-memory build (53 MB download, ~120 MB of memory)")
+    p.add_argument("--model", default=DEFAULT_REPO, help="Hugging Face repo id or local folder")
+    p.add_argument("--force-language", help="skip detection: the model's own language or en")
+    p.add_argument("--numbers", choices=["words", "digits"], default="words")
+    p.add_argument("--greedy", action="store_true", help="faster, slightly less accurate")
+    p.add_argument("--threads", type=int, default=2)
+    a = p.parse_args(argv)
+    try:
+        asr = Shruti.load(a.lang, "int8" if a.int8 else "4bit", a.model, a.threads)
+    except ValueError as e:
+        p.error(str(e))
+    for f in a.audio:
+        r = asr.transcribe(f, language=a.force_language, beams=1 if a.greedy else 5, numbers=a.numbers)
+        print(f"{f}\t{r['language']}\t{r['text']}")
+
+
+if __name__ == "__main__":
+    main()

shruti_onnx-0.2.0/shruti_onnx/__main__.py

@@ -0,0 +1,3 @@
+from . import main
+
+main()

shruti_onnx-0.2.0/shruti_onnx.egg-info/PKG-INFO

@@ -0,0 +1,46 @@
+Metadata-Version: 2.4
+Name: shruti-onnx
+Version: 0.2.0
+Summary: Shruti: offline speech-to-text, one language and English in one small model, in ONNX Runtime (32 MB).
+Author: Udgam Labs
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://huggingface.co/udgamlabs/Shruti-ONNX
+Project-URL: Source, https://github.com/udgamlabs/shruti
+Keywords: speech-recognition,speech-to-text,onnx,offline,on-device
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy
+Requires-Dist: onnxruntime>=1.17
+Requires-Dist: huggingface_hub
+Provides-Extra: audio
+Requires-Dist: soundfile; extra == "audio"
+Requires-Dist: soxr; extra == "audio"
+Dynamic: license-file
+
+# shruti-onnx
+
+Shruti: offline speech-to-text, one small model per language. Each model covers its language, English, and the two mixed. Models and full results: [udgamlabs/Shruti-ONNX](https://huggingface.co/udgamlabs/Shruti-ONNX).
+
+Languages: Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu, Urdu. Use the word or the code.
+
+```bash
+pip install shruti-onnx
+```
+
+```python
+from shruti_onnx import Shruti
+
+asr = Shruti.load("urdu")                    # or "ur"; variant="int8" for low memory on ARM
+print(asr.transcribe("clip.wav")["text"])    # language detected; language="en" to set English
+```
+
+Command line: `shruti-onnx clip.wav --lang urdu`. `loops="cut"` removes a repeat and what follows it.
+
+## Where it falls short
+
+Nine of these languages (Dogri, Hindi, Kannada, Konkani, Odia, Punjabi, Sindhi, Tamil, Telugu) failed our written comparison against the original engine, and none met every rule. The model card names each miss and its size. On Linux x86, use the default 4-bit build: the int8 and sherpa-onnx builds were less accurate there on the machine we tested. A model can get stuck repeating a phrase; the text stays as the model wrote it unless you ask to cut it. The 4-bit build needs about 1 GB of memory once loaded in ONNX Runtime.
+
+Apache-2.0. Built on [Cactus Whistle](https://huggingface.co/Cactus-Compute/whistle).
+
+**About Udgam Labs.** Udgam Labs is an independent AI research lab. [udgamlabs.com](https://udgamlabs.com)

shruti_onnx-0.2.0/shruti_onnx.egg-info/SOURCES.txt

@@ -0,0 +1,12 @@
+LICENSE
+README.md
+pyproject.toml
+shruti_onnx/__init__.py
+shruti_onnx/__main__.py
+shruti_onnx.egg-info/PKG-INFO
+shruti_onnx.egg-info/SOURCES.txt
+shruti_onnx.egg-info/dependency_links.txt
+shruti_onnx.egg-info/entry_points.txt
+shruti_onnx.egg-info/requires.txt
+shruti_onnx.egg-info/top_level.txt
+tests/test_same_as_research_runner.py

shruti_onnx-0.2.0/shruti_onnx.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

shruti_onnx-0.2.0/shruti_onnx.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+shruti-onnx = shruti_onnx:main

shruti_onnx-0.2.0/shruti_onnx.egg-info/requires.txt

@@ -0,0 +1,7 @@
+numpy
+onnxruntime>=1.17
+huggingface_hub
+
+[audio]
+soundfile
+soxr

shruti_onnx-0.2.0/shruti_onnx.egg-info/top_level.txt

@@ -0,0 +1 @@
+shruti_onnx

shruti_onnx-0.2.0/tests/test_same_as_research_runner.py

@@ -0,0 +1,36 @@
+"""The package must transcribe exactly like the Gate 0 research runner (udgam-voice/asr/onnx/shruti_onnx.py), with
+its frozen decoder settings (beam 5, length-normalised), on the phone-test clips; plus file reading and digits.
+Run (onnx venv): python tests/test_same_as_research_runner.py <model_dir> [n_clips]"""
+import importlib.util
+import os
+import sys
+
+import soundfile as sf
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+TRAIN = os.path.abspath(os.path.join(HERE, "../../../.."))
+sys.path.insert(0, os.path.join(HERE, ".."))
+from shruti_onnx import Shruti  # noqa: E402
+
+spec = importlib.util.spec_from_file_location("research", os.path.join(TRAIN, "udgam-voice/asr/onnx/shruti_onnx.py"))
+research = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(research)
+
+model, n = sys.argv[1], int(sys.argv[2]) if len(sys.argv) > 2 else 20
+pkg, ref = Shruti(model), research.Shruti(model, threads=2)
+clips = os.path.join(TRAIN, "udgam-voice/asr/work/phone_test/clips")
+same = 0
+names = sorted(os.listdir(clips))[:n]
+for c in names:
+    path = os.path.join(clips, c)
+    w, _ = sf.read(path, dtype="float32")
+    a = pkg.transcribe(path)["text"]                       # from the file, language detected
+    b = ref.transcribe(w, language=None, beams=5, length_norm=True)["text"]
+    same += a == b
+    if a != b:
+        print("DIFF", c, "\n pkg:", a, "\n ref:", b)
+print(f"package == research runner on {same}/{len(names)} clips")
+sys.path.insert(0, os.path.join(TRAIN, "textnorm/v1/python"))
+r = pkg.transcribe(os.path.join(clips, "fl08.wav"), numbers="digits")
+print("digits:", r["language"], r["text"])
+assert same == len(names)