afrospeech-recipes 0.1.0__tar.gz

afrospeech_recipes-0.1.0/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 2026 AfriModal AI
+
+   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.

afrospeech_recipes-0.1.0/PKG-INFO

@@ -0,0 +1,102 @@
+Metadata-Version: 2.4
+Name: afrospeech-recipes
+Version: 0.1.0
+Summary: Load the speech data a recipe describes, from wherever it lives.
+Author-email: AfriModal AI <kaykola.olaleye@gmail.com>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://afrimodal-ai.github.io/
+Keywords: speech,asr,african-languages,datasets,streaming
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Multimedia :: Sound/Audio :: Speech
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: requests>=2.28
+Requires-Dist: numpy>=1.23
+Requires-Dist: soundfile>=0.12
+Requires-Dist: scipy>=1.10
+Provides-Extra: huggingface
+Requires-Dist: datasets>=2.18; extra == "huggingface"
+Dynamic: license-file
+
+# The recipe loader
+
+One call that loads exactly the speech a recipe describes, wherever the data
+lives. The recipe comes from the builder page; paste it into a notebook and
+run it.
+
+```python
+from afrospeech_recipes import load_recipe
+
+ds = load_recipe(recipe, split="train", sample_rate=16_000, stream=True)
+for ex in ds.take(3):
+    print(ex["source"], round(ex["duration"], 1), "s", ex["text"])
+print(ds.report())
+```
+
+Install it:
+
+```
+pip install "afrospeech-recipes[huggingface]==0.1.0"
+```
+
+The extra brings `datasets`, which only a Hugging Face source needs; a
+recipe that reads a tar archive runs without it. Every script the builder
+page hands out pins the version it was written against, because a recipe is
+a record of what was fetched and a loader that changed under it is a
+different answer.
+
+## The data keeps its own terms
+
+This licence covers the loader, not one second of audio. Every dataset keeps
+the terms its publisher set: attribution, share-alike, no redistribution, an
+access request, a non-commercial clause. A recipe carries the permission
+answers that were recorded for each source and this loader checks them again
+before it fetches, but a check in code is not a licence review, and a
+permission nobody has read is not permission. Read each publisher's terms
+before you train on their audio, and before you publish anything you trained.
+
+## What version 1 does
+
+- Hugging Face repositories, streamed: parquet, webdataset and audio folders.
+- Tar archives over HTTP, read as a stream and never written to disk.
+- Filters whole datasets, which is what the catalogue records today.
+- Shares an hour budget over the sources using each dataset's recorded
+  hours, so a budget is met to within one clip and `report()` says what was
+  actually delivered.
+- Resamples to one rate and returns one record shape whatever the source:
+  `audio, text, duration, language, split, source, clip_id, speaker_id`.
+
+## What it refuses to do
+
+- Mix splits. A training recipe reads train splits only, and asking it for a
+  test split raises rather than returning the audio.
+- Read a source whose recorded terms do not allow the recipe's purpose. The
+  permission answers travel inside the recipe and are checked again here, at
+  the point the data is fetched.
+- Keep audio. Clips are decoded, handed over and dropped.
+
+## What is not here yet
+
+- Per clip selection: length, speaker, topic and an exact hour budget all
+  need the clip index, which is the next step.
+- Adapters for zip archives, GitHub releases, cloud buckets, Dataverse and
+  Zenodo, and for local copies of datasets that need a signed agreement.
+- Writing shards or a NeMo manifest instead of streaming.
+
+A dataset this loader cannot yet fetch correctly is never written into a
+script. The builder page counts its hours, names it under the script, and
+says why it is not in the code.
+
+## Which datasets it can fetch
+
+The loader fetches whatever a recipe names, and a recipe is written by the
+builder page, which only writes a source it has checked: it asks each
+publisher's repository what configs and splits it really has and which
+columns a row carries, and records the date it asked. Nothing there is
+inferred from a naming convention, and a dataset whose train and test split
+is recorded somewhere other than the data itself is left out with that as
+the stated reason, rather than loaded on a guess.

afrospeech_recipes-0.1.0/README.md

@@ -0,0 +1,78 @@
+# The recipe loader
+
+One call that loads exactly the speech a recipe describes, wherever the data
+lives. The recipe comes from the builder page; paste it into a notebook and
+run it.
+
+```python
+from afrospeech_recipes import load_recipe
+
+ds = load_recipe(recipe, split="train", sample_rate=16_000, stream=True)
+for ex in ds.take(3):
+    print(ex["source"], round(ex["duration"], 1), "s", ex["text"])
+print(ds.report())
+```
+
+Install it:
+
+```
+pip install "afrospeech-recipes[huggingface]==0.1.0"
+```
+
+The extra brings `datasets`, which only a Hugging Face source needs; a
+recipe that reads a tar archive runs without it. Every script the builder
+page hands out pins the version it was written against, because a recipe is
+a record of what was fetched and a loader that changed under it is a
+different answer.
+
+## The data keeps its own terms
+
+This licence covers the loader, not one second of audio. Every dataset keeps
+the terms its publisher set: attribution, share-alike, no redistribution, an
+access request, a non-commercial clause. A recipe carries the permission
+answers that were recorded for each source and this loader checks them again
+before it fetches, but a check in code is not a licence review, and a
+permission nobody has read is not permission. Read each publisher's terms
+before you train on their audio, and before you publish anything you trained.
+
+## What version 1 does
+
+- Hugging Face repositories, streamed: parquet, webdataset and audio folders.
+- Tar archives over HTTP, read as a stream and never written to disk.
+- Filters whole datasets, which is what the catalogue records today.
+- Shares an hour budget over the sources using each dataset's recorded
+  hours, so a budget is met to within one clip and `report()` says what was
+  actually delivered.
+- Resamples to one rate and returns one record shape whatever the source:
+  `audio, text, duration, language, split, source, clip_id, speaker_id`.
+
+## What it refuses to do
+
+- Mix splits. A training recipe reads train splits only, and asking it for a
+  test split raises rather than returning the audio.
+- Read a source whose recorded terms do not allow the recipe's purpose. The
+  permission answers travel inside the recipe and are checked again here, at
+  the point the data is fetched.
+- Keep audio. Clips are decoded, handed over and dropped.
+
+## What is not here yet
+
+- Per clip selection: length, speaker, topic and an exact hour budget all
+  need the clip index, which is the next step.
+- Adapters for zip archives, GitHub releases, cloud buckets, Dataverse and
+  Zenodo, and for local copies of datasets that need a signed agreement.
+- Writing shards or a NeMo manifest instead of streaming.
+
+A dataset this loader cannot yet fetch correctly is never written into a
+script. The builder page counts its hours, names it under the script, and
+says why it is not in the code.
+
+## Which datasets it can fetch
+
+The loader fetches whatever a recipe names, and a recipe is written by the
+builder page, which only writes a source it has checked: it asks each
+publisher's repository what configs and splits it really has and which
+columns a row carries, and records the date it asked. Nothing there is
+inferred from a naming convention, and a dataset whose train and test split
+is recorded somewhere other than the data itself is left out with that as
+the stated reason, rather than loaded on a guess.

afrospeech_recipes-0.1.0/afrospeech_recipes/__init__.py

@@ -0,0 +1,251 @@
+"""Load exactly the speech data a recipe describes, from wherever it lives.
+
+    from afrospeech_recipes import load_recipe
+
+    ds = load_recipe(recipe, split="train", sample_rate=16_000, stream=True)
+    print(ds.plan())          # what it will ask each publisher for
+    for ex in ds.take(3):
+        print(ex["source"], round(ex["duration"], 1), "s", ex["text"])
+    print(ds.report())        # what actually arrived
+
+One call, many sources underneath. The recipe comes from the builder page and
+carries everything needed, so this package reads no catalogue and contacts no
+site of ours: only the dataset publishers themselves.
+
+Version 1 handles Hugging Face repositories and tar archives, filters whole
+datasets, and shares an hour budget using each dataset's recorded hours, so a
+budget is met to within a clip rather than exactly. Per clip filters (length,
+speaker, topic) and exact budgets arrive with the clip index.
+
+What it will not do:
+
+  - Mix splits. A training recipe reads train splits only, and asking it for
+    a test split raises instead.
+  - Load a source whose recorded terms do not allow the recipe's purpose,
+    even if the recipe lists it. The answers travel with the recipe and are
+    checked again here, where the fetching happens.
+  - Keep audio. Clips are decoded, handed over and dropped. Nothing is
+    written to disk unless `stream=False`, which turns on the publisher's own
+    caching.
+"""
+
+from __future__ import annotations
+
+import itertools
+from typing import Iterator
+
+from .recipe import Recipe, RecipeError, Source, budget, parse, role_for, usable
+from .sources import adapter
+
+__all__ = ["load_recipe", "RecipeStream", "RecipeError", "Recipe", "Source"]
+
+DEFAULT_SAMPLE_RATE = 16000
+
+
+def span(hours: float) -> str:
+    """An amount of audio, in the unit it is actually in.
+
+    A three clip check delivers six seconds, and "0.0 h" reads as nothing
+    delivered when three clips did arrive. Below a tenth of an hour this says
+    seconds instead.
+    """
+    if hours >= 0.1:
+        return f"{hours:.1f} h"
+    return f"{hours * 3600:.0f} s"
+
+
+class RecipeStream:
+    """The clips a recipe asks for, in a mix of its sources.
+
+    Iterable once per pass, like any stream. `take(n)` is the short form the
+    site's snippet uses. `report()` says what was actually delivered, which
+    is the only honest way to state an approximate budget.
+    """
+
+    def __init__(self, recipe: Recipe, role: str, sample_rate: int, streaming: bool,
+                 on_source_error: str = "stop"):
+        self.recipe = recipe
+        self.role = role
+        self.sample_rate = sample_rate
+        self.streaming = streaming
+        self.on_source_error = on_source_error
+        self.delivered: dict[str, float] = {}
+        self.clips: dict[str, int] = {}
+        self.skipped: dict[str, str] = {}
+        self.failed: dict[str, str] = {}
+
+        self.sources: list[Source] = []
+        for source in recipe.sources:
+            why = usable(recipe, source, role)
+            if why:
+                self.skipped[f"{source.name} ({source.language})"] = why
+            else:
+                self.sources.append(source)
+        if not self.sources:
+            raise RecipeError(
+                "no source in this recipe can be read for "
+                f"{recipe.purpose}: " + "; ".join(
+                    f"{k}: {v}" for k, v in self.skipped.items()))
+        self.budget = budget(recipe, self.sources)
+
+    # -- the records -------------------------------------------------------
+
+    def _labelled_out(self, source: Source, row: dict) -> bool:
+        """Whether this clip's own labels put it outside the recipe.
+
+        A source whose publisher records no such column cannot be filtered
+        clip by clip: the site has already decided whether to include it
+        whole, so nothing is dropped here on a guess.
+        """
+        for field_name in ("speech_style", "topic"):
+            column = source.labels.get(field_name)
+            if not column:
+                continue
+            value = row.get(column)
+            value = None if value is None else str(value).strip()
+            if not value:
+                if field_name in self.recipe.require_labels:
+                    return True
+                continue
+            for unwanted in self.recipe.excluded(field_name):
+                if value.casefold() == str(unwanted).casefold():
+                    return True
+        return False
+
+    def _record(self, source: Source, row: dict) -> dict | None:
+        if self._labelled_out(source, row):
+            return None
+        audio = row.get(source.audio)
+        if not isinstance(audio, dict) or audio.get("array") is None:
+            return None
+        rate = audio.get("sampling_rate") or self.sample_rate
+        seconds = len(audio["array"]) / rate if rate else None
+        if source.duration and row.get(source.duration) is not None:
+            try:
+                seconds = float(row[source.duration])
+            except (TypeError, ValueError):
+                pass
+        if seconds is None:
+            return None
+        if self.recipe.max_clip_seconds and seconds > self.recipe.max_clip_seconds:
+            return None
+        text = row.get(source.text)
+        if text is None or str(text).strip() == "":
+            return None
+        return {
+            "audio": audio,
+            "text": str(text),
+            "duration": seconds,
+            "language": source.language,
+            "split": self.role,
+            "source": source.name,
+            "clip_id": audio.get("path") or row.get("id"),
+            "speaker_id": row.get(source.speaker) if source.speaker else None,
+        }
+
+    def _one(self, source: Source) -> Iterator[dict]:
+        """One source's clips, stopping at its share of the budget."""
+        name = f"{source.name}/{source.language}"
+        allowance = self.budget.get(name)
+        taken = 0.0
+        try:
+            rows = adapter(source.loader)(source, self.role, self.sample_rate,
+                                          self.streaming)
+            for row in rows:
+                record = self._record(source, row)
+                if record is None:
+                    continue
+                if allowance is not None and taken + record["duration"] > allowance * 3600:
+                    break
+                taken += record["duration"]
+                self.delivered[name] = taken / 3600
+                self.clips[name] = self.clips.get(name, 0) + 1
+                yield record
+        except Exception as exc:  # one source being down is not the run's end
+            self.failed[name] = f"{type(exc).__name__}: {exc}"
+            if self.on_source_error == "stop":
+                raise RuntimeError(
+                    f"{name} could not be read: {exc}. Pass "
+                    "on_source_error='skip' to carry on without it.") from exc
+
+    def __iter__(self) -> Iterator[dict]:
+        """Sources interleaved, one clip at a time round the mix, so a short
+        run holds the same mixture as a long one."""
+        streams = [self._one(source) for source in self.sources]
+        while streams:
+            for stream in list(streams):
+                try:
+                    yield next(stream)
+                except StopIteration:
+                    streams.remove(stream)
+
+    def take(self, n: int) -> Iterator[dict]:
+        return itertools.islice(iter(self), n)
+
+    # -- what happened -----------------------------------------------------
+
+    def plan(self) -> str:
+        """What this stream will ask each publisher for, before it asks.
+
+        Printed by the script the site hands out, because the hours on the
+        page are a plan and this is the same plan as the code holds it: the
+        share per source, the split each share is read from, and the sources
+        the recipe named that this run will not read.
+        """
+        asked = 0.0
+        for hours in self.budget.values():
+            asked += hours or 0.0
+        total = f"{asked:.0f} h" if asked else "every hour on offer"
+        lines = [f"Asking for {total} of {'/'.join(self.recipe.languages)} "
+                 f"for {self.recipe.purpose}, from the {self.role} split of "
+                 f"{len(self.sources)} source"
+                 f"{'' if len(self.sources) == 1 else 's'}:"]
+        for source in self.sources:
+            name = f"{source.name}/{source.language}"
+            share = self.budget.get(name)
+            lines.append(f"  {name}: "
+                         + (f"up to {share:.1f} h" if share else "everything")
+                         + f', split "{source.split_name(self.role)}"')
+        for name, why in sorted(self.skipped.items()):
+            lines.append(f"  not read: {name}: {why}")
+        if self.recipe.max_clip_seconds:
+            lines.append(f"  clips longer than {self.recipe.max_clip_seconds:g} s "
+                         "are dropped as they arrive")
+        return "\n".join(lines)
+
+    def report(self) -> str:
+        total = 0.0
+        for hours in self.delivered.values():
+            total += hours
+        clips = 0
+        for count in self.clips.values():
+            clips += count
+        lines = [f"Delivered {span(total)} in {clips} clip{'' if clips == 1 else 's'} from "
+                 f"{len(self.delivered)} source{'' if len(self.delivered) == 1 else 's'}, "
+                 "measured from the audio as it streamed."]
+        for name in sorted(self.delivered):
+            share = self.budget.get(name)
+            count = self.clips.get(name, 0)
+            lines.append(f"  {name}: {span(self.delivered[name])} "
+                         f"in {count} clip{'' if count == 1 else 's'}"
+                         + (f" of up to {share:.1f} h asked for" if share else ""))
+        for name, why in sorted(self.skipped.items()):
+            lines.append(f"  not read: {name}: {why}")
+        for name, why in sorted(self.failed.items()):
+            lines.append(f"  failed: {name}: {why}")
+        return "\n".join(lines)
+
+
+def load_recipe(recipe, split: str | None = None,
+                sample_rate: int = DEFAULT_SAMPLE_RATE, stream: bool = True,
+                on_source_error: str = "stop") -> RecipeStream:
+    """Open a recipe as a stream of clips.
+
+    `split` names the role to read, and must agree with the recipe's purpose;
+    left out, the purpose decides. `stream=False` lets the publisher's own
+    library cache to disk, which is for a machine with room, never for a free
+    tier.
+    """
+    parsed = parse(recipe)
+    return RecipeStream(parsed, role_for(parsed, split), sample_rate, stream,
+                        on_source_error)

afrospeech_recipes-0.1.0/afrospeech_recipes/recipe.py

@@ -0,0 +1,184 @@
+"""A recipe, checked before anything is fetched.
+
+A recipe is a plain dict, written by the builder page and pasted into a
+notebook. It carries everything the loader needs, so the loader asks no
+website anything: no catalogue lookup, no index download, no account.
+
+The rules are kept here rather than in the adapters, because every one of
+them is about the recipe as a whole: a split role a source does not publish is refused, a source whose
+recorded answer does not allow the recipe's purpose is refused, and a budget
+is shared over the sources before a single clip is read.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+# What a recipe may ask for, and the split role each one reads.
+PURPOSES = {
+    "asr_training": "train",
+    "asr_evaluation": "test",
+    # Quantising audio into discrete units for language model pretraining.
+    # Deliberately its own answer and never a synonym for asr_training:
+    # config/licences.yaml says why, and a corpus that forbids voice cloning
+    # may well forbid this while allowing recognition.
+    "speech_lm_pretraining": "train",
+    "tts_training": "train",
+}
+USES = ("research", "commercial")
+ROLES = ("train", "dev", "test")
+
+
+class RecipeError(ValueError):
+    """A recipe that cannot be run, said plainly enough to fix."""
+
+
+@dataclass
+class Source:
+    """One dataset, in one language, as the recipe names it."""
+
+    name: str
+    language: str
+    repo: str
+    splits: dict[str, str]
+    revision: str | None = None
+    where: dict | None = None
+    hours: float | None = None
+    audio: str = "audio"
+    text: str = "text"
+    duration: str | None = None
+    speaker: str | None = None
+    allows: list[str] = field(default_factory=list)
+    loader: str = "huggingface"
+    url: str | None = None
+    # Which of this repository's own columns says what a clip is: the same
+    # columns the hours on the site were counted from.
+    labels: dict[str, str] = field(default_factory=dict)
+
+    def split_name(self, role: str) -> str | None:
+        return self.splits.get(role)
+
+
+@dataclass
+class Recipe:
+    languages: list[str]
+    purpose: str
+    use: str
+    sources: list[Source]
+    hours: float | None = None
+    seed: int = 42
+    max_clip_seconds: float | None = None
+    # Values the visitor turned off, as the publishers spell them, and the
+    # fields whose unlabelled clips they also turned off.
+    exclude_speech_style: list[str] = field(default_factory=list)
+    exclude_topics: list[str] = field(default_factory=list)
+    require_labels: list[str] = field(default_factory=list)
+    split_hours: str = "proportional"
+    recipe_version: int = 1
+
+    def excluded(self, field_name: str) -> list[str]:
+        return (self.exclude_speech_style if field_name == "speech_style"
+                else self.exclude_topics)
+
+
+def parse(recipe: dict) -> Recipe:
+    if not isinstance(recipe, dict):
+        raise RecipeError("a recipe is a dict, as the site writes it")
+    for required in ("languages", "purpose", "use", "sources"):
+        if required not in recipe:
+            raise RecipeError(f"the recipe has no {required!r}")
+    purpose = recipe["purpose"]
+    if purpose not in PURPOSES:
+        raise RecipeError(f"purpose {purpose!r} is not one of {sorted(PURPOSES)}")
+    if recipe["use"] not in USES:
+        raise RecipeError(f"use {recipe['use']!r} is not one of {list(USES)}")
+    if not recipe["sources"]:
+        raise RecipeError("the recipe has no sources, so there is nothing to load")
+
+    sources = []
+    for entry in recipe["sources"]:
+        known = {f: entry[f] for f in Source.__dataclass_fields__ if f in entry}
+        for required in ("name", "language", "repo", "splits"):
+            if required not in known:
+                raise RecipeError(f"a source has no {required!r}: {entry.get('name', entry)}")
+        sources.append(Source(**known))
+
+    return Recipe(
+        languages=list(recipe["languages"]),
+        purpose=purpose,
+        use=recipe["use"],
+        sources=sources,
+        hours=recipe.get("hours"),
+        seed=recipe.get("seed", 42),
+        max_clip_seconds=recipe.get("max_clip_seconds"),
+        exclude_speech_style=list(recipe.get("exclude_speech_style") or []),
+        exclude_topics=list(recipe.get("exclude_topics") or []),
+        require_labels=list(recipe.get("require_labels") or []),
+        split_hours=recipe.get("split_hours", "proportional"),
+        recipe_version=recipe.get("recipe_version", 1),
+    )
+
+
+def role_for(recipe: Recipe, split: str | None) -> str:
+    """Which split role this run reads.
+
+    `split` may be given explicitly, but it must agree with the purpose: a
+    training recipe never reads a test split, whatever the caller passes
+    Train and test never mix.
+    """
+    wanted = PURPOSES[recipe.purpose]
+    if split is None:
+        return wanted
+    if split not in ROLES:
+        raise RecipeError(f"split {split!r} is not one of {list(ROLES)}")
+    if split == "test" and wanted != "test":
+        raise RecipeError(
+            f"this recipe's purpose is {recipe.purpose}, which trains on the train "
+            "split. Loading the test split for it would train on test audio, so the "
+            "loader refuses. Build an evaluation recipe instead.")
+    if split != "test" and wanted == "test":
+        raise RecipeError(
+            f"this recipe's purpose is {recipe.purpose}, which reads the test split.")
+    return split
+
+
+def usable(recipe: Recipe, source: Source, role: str) -> str | None:
+    """Why this source cannot be read for this recipe, or None if it can.
+
+    The permission answers the site recorded travel with the recipe, so the
+    check happens again here, at run time, where the data is actually
+    fetched.
+    """
+    if source.split_name(role) is None:
+        return f"publishes no {role} split"
+    if source.allows:
+        if recipe.purpose not in source.allows:
+            return f"its recorded terms do not allow {recipe.purpose}"
+        if recipe.use == "commercial" and "commercial" not in source.allows:
+            return "its recorded terms do not allow commercial use"
+    return None
+
+
+def budget(recipe: Recipe, sources: list[Source]) -> dict[str, float | None]:
+    """How many hours to take from each source, keyed by name and language.
+
+    The site already shared the budget out and wrote each share into the
+    recipe, so that is what is used. Without one, the recipe's total is
+    shared in proportion to what each source offers, and a recipe with no
+    total at all takes everything.
+    """
+    out: dict[str, float | None] = {}
+    stated = [s for s in sources if s.hours]
+    if len(stated) == len(sources) and stated:
+        for source in sources:
+            out[f"{source.name}/{source.language}"] = source.hours
+        return out
+    if recipe.hours is None:
+        return {f"{s.name}/{s.language}": None for s in sources}
+    offered = 0.0
+    for source in sources:
+        offered += source.hours or 0.0
+    for source in sources:
+        share = (source.hours or 0.0) / offered if offered else 1 / len(sources)
+        out[f"{source.name}/{source.language}"] = recipe.hours * share
+    return out

afrospeech_recipes-0.1.0/afrospeech_recipes/sources/__init__.py

@@ -0,0 +1,20 @@
+"""One adapter per kind of source. Each one yields the repository's own rows;
+turning a row into the common record is the loader's job, not theirs."""
+
+from __future__ import annotations
+
+from . import huggingface, tarball
+
+ADAPTERS = {
+    "huggingface": huggingface.stream,
+    "tar": tarball.stream,
+}
+
+
+def adapter(name: str):
+    if name not in ADAPTERS:
+        raise RuntimeError(
+            f"no adapter for a {name!r} source. The loader can fetch "
+            f"{sorted(ADAPTERS)} so far; the site never writes a recipe that "
+            "names anything else.")
+    return ADAPTERS[name]

afrospeech_recipes-0.1.0/afrospeech_recipes/sources/audio.py

@@ -0,0 +1,53 @@
+"""Turning a clip's bytes into samples, once, for every adapter.
+
+Decoding is done here rather than left to the publisher's own library on
+purpose. The `datasets` library decodes audio through torchcodec, which
+means installing torch to read a wav file; on a free tier that is most of
+the disk allowance spent before any data arrives. soundfile reads what the
+catalogued datasets actually ship, and scipy resamples, and both are small.
+"""
+
+from __future__ import annotations
+
+import io
+from math import gcd
+
+
+def decode(payload: bytes, sample_rate: int):
+    """One clip's samples at the rate asked for, mono, as float32."""
+    import numpy
+    import soundfile
+
+    array, rate = soundfile.read(io.BytesIO(payload), dtype="float32", always_2d=True)
+    array = array.mean(axis=1)
+    if rate != sample_rate:
+        from scipy.signal import resample_poly
+
+        divisor = gcd(int(rate), int(sample_rate))
+        array = resample_poly(array, int(sample_rate) // divisor, int(rate) // divisor)
+    return numpy.asarray(array, dtype="float32"), sample_rate
+
+
+def as_clip(value, sample_rate: int) -> dict | None:
+    """An audio column's value, however the source gives it, as one clip.
+
+    A row may carry the bytes (what a parquet repository holds), a path to a
+    file, or samples a library has already decoded. Anything else is not
+    audio this adapter understands, and the caller reports it rather than
+    guessing.
+    """
+    if not isinstance(value, dict):
+        return None
+    if value.get("array") is not None:
+        return value
+    payload = value.get("bytes")
+    if payload is None and value.get("path"):
+        try:
+            with open(value["path"], "rb") as handle:
+                payload = handle.read()
+        except OSError:
+            return None
+    if payload is None:
+        return None
+    array, rate = decode(payload, sample_rate)
+    return {"array": array, "sampling_rate": rate, "path": value.get("path")}

afrospeech_recipes-0.1.0/afrospeech_recipes/sources/huggingface.py

@@ -0,0 +1,67 @@
+"""Hugging Face: parquet, webdataset and audio folders, streamed.
+
+The one thing this adapter is careful about is which rows belong to the
+split the recipe asked for. A repo says that in one of two ways, and the
+recipe says which: the split is a split the repo publishes, or the language
+is a config the repo publishes and the split is inside it. Anything else
+(a split recorded in a file beside the data, say) is not something this
+adapter can tell apart, so the site does not offer those datasets yet
+rather than guessing and mixing test audio into a training run.
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Any, Iterator
+
+from .audio import as_clip
+
+
+def _datasets():
+    try:
+        import datasets
+    except ImportError as exc:  # pragma: no cover - a missing dependency
+        raise RuntimeError(
+            "this recipe loads a Hugging Face dataset, so it needs the "
+            "`datasets` package: pip install datasets") from exc
+    return datasets
+
+
+def stream(source, role: str, sample_rate: int, streaming: bool = True) -> Iterator[dict]:
+    """Yield one dict per clip, as the repo stores it, audio decoded and
+    resampled. Rows the recipe's language filter rejects are skipped here,
+    so the caller only ever sees the language it asked for."""
+    datasets = _datasets()
+    where = source.where or {}
+    split = source.split_name(role)
+    kwargs: dict[str, Any] = {"split": split, "streaming": streaming}
+    if source.revision:
+        kwargs["revision"] = source.revision
+    token = os.environ.get("HF_TOKEN")
+    if token:
+        kwargs["token"] = token
+
+    if where.get("config"):
+        dataset = datasets.load_dataset(source.repo, where["config"], **kwargs)
+    elif where.get("data_dir"):
+        dataset = datasets.load_dataset(source.repo, data_dir=where["data_dir"], **kwargs)
+    else:
+        dataset = datasets.load_dataset(source.repo, **kwargs)
+
+    # Hand back the clip's bytes rather than letting `datasets` decode them.
+    # Its decoder is torchcodec, which means installing torch to read a wav
+    # file: most of a free tier's disk spent before any audio arrives. The
+    # bytes are decoded in sources/audio.py with soundfile instead.
+    try:
+        dataset = dataset.cast_column(source.audio, datasets.Audio(decode=False))
+    except (ValueError, TypeError):
+        pass
+
+    column, value = where.get("column"), where.get("value")
+    for row in dataset:
+        if column and row.get(column) != value:
+            continue
+        clip = as_clip(row.get(source.audio), sample_rate)
+        if clip is None:
+            continue
+        yield dict(row, **{source.audio: clip})

afrospeech_recipes-0.1.0/afrospeech_recipes/sources/tarball.py

@@ -0,0 +1,69 @@
+"""A .tar or .tar.gz of audio, read as a stream.
+
+OpenSLR and several university mirrors publish a corpus as one archive. The
+honest limit, which the site states next to any dataset that uses this
+adapter: the whole archive travels down the wire even when the recipe wants
+a small part of it, because a gzipped tar cannot be read out of order. The
+archive is never written to disk. Members are read one at a time and
+dropped as soon as the clip has been handed over.
+
+Transcripts come from a separate index the recipe points at, mapping a
+member path to its text. An archive member with no entry in that index is
+skipped rather than returned with an empty transcript.
+"""
+
+from __future__ import annotations
+
+import tarfile
+from typing import Iterator
+
+from .audio import decode
+
+
+def _open_stream(url: str, token: str | None = None):
+    import requests
+
+    headers = {"Authorization": f"Bearer {token}"} if token else {}
+    response = requests.get(url, stream=True, timeout=120, headers=headers)
+    response.raise_for_status()
+    response.raw.decode_content = False
+    return response.raw
+
+
+def members(url: str, suffixes=(".wav", ".flac", ".mp3", ".ogg"),
+            token: str | None = None, fileobj=None) -> Iterator[tuple[str, bytes]]:
+    """Every audio member of the archive, in the order it was written.
+
+    `fileobj` is for tests and for a local archive: anything the tarfile
+    module can read in stream mode. `r|*` is the streaming mode, which never
+    seeks, so it works on a socket.
+    """
+    handle = fileobj if fileobj is not None else _open_stream(url, token)
+    with tarfile.open(fileobj=handle, mode="r|*") as archive:
+        for member in archive:
+            if not member.isfile():
+                continue
+            if suffixes and not member.name.lower().endswith(tuple(suffixes)):
+                continue
+            extracted = archive.extractfile(member)
+            if extracted is None:
+                continue
+            yield member.name, extracted.read()
+
+
+def stream(source, role: str, sample_rate: int, streaming: bool = True) -> Iterator[dict]:
+    """Yield one row per archive member that the transcript index names."""
+    import os
+
+    index = (source.where or {}).get("transcripts") or {}
+    url = source.url or (source.where or {}).get("url")
+    if not url:
+        raise RuntimeError(f"{source.name}: the recipe gives no archive URL")
+    token = os.environ.get("HF_TOKEN")
+    for name, payload in members(url, token=token):
+        text = index.get(name)
+        if text is None:
+            continue
+        array, rate = decode(payload, sample_rate)
+        yield {source.audio: {"array": array, "sampling_rate": rate, "path": name},
+               source.text: text}

afrospeech_recipes-0.1.0/afrospeech_recipes.egg-info/PKG-INFO

@@ -0,0 +1,102 @@
+Metadata-Version: 2.4
+Name: afrospeech-recipes
+Version: 0.1.0
+Summary: Load the speech data a recipe describes, from wherever it lives.
+Author-email: AfriModal AI <kaykola.olaleye@gmail.com>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://afrimodal-ai.github.io/
+Keywords: speech,asr,african-languages,datasets,streaming
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Multimedia :: Sound/Audio :: Speech
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: requests>=2.28
+Requires-Dist: numpy>=1.23
+Requires-Dist: soundfile>=0.12
+Requires-Dist: scipy>=1.10
+Provides-Extra: huggingface
+Requires-Dist: datasets>=2.18; extra == "huggingface"
+Dynamic: license-file
+
+# The recipe loader
+
+One call that loads exactly the speech a recipe describes, wherever the data
+lives. The recipe comes from the builder page; paste it into a notebook and
+run it.
+
+```python
+from afrospeech_recipes import load_recipe
+
+ds = load_recipe(recipe, split="train", sample_rate=16_000, stream=True)
+for ex in ds.take(3):
+    print(ex["source"], round(ex["duration"], 1), "s", ex["text"])
+print(ds.report())
+```
+
+Install it:
+
+```
+pip install "afrospeech-recipes[huggingface]==0.1.0"
+```
+
+The extra brings `datasets`, which only a Hugging Face source needs; a
+recipe that reads a tar archive runs without it. Every script the builder
+page hands out pins the version it was written against, because a recipe is
+a record of what was fetched and a loader that changed under it is a
+different answer.
+
+## The data keeps its own terms
+
+This licence covers the loader, not one second of audio. Every dataset keeps
+the terms its publisher set: attribution, share-alike, no redistribution, an
+access request, a non-commercial clause. A recipe carries the permission
+answers that were recorded for each source and this loader checks them again
+before it fetches, but a check in code is not a licence review, and a
+permission nobody has read is not permission. Read each publisher's terms
+before you train on their audio, and before you publish anything you trained.
+
+## What version 1 does
+
+- Hugging Face repositories, streamed: parquet, webdataset and audio folders.
+- Tar archives over HTTP, read as a stream and never written to disk.
+- Filters whole datasets, which is what the catalogue records today.
+- Shares an hour budget over the sources using each dataset's recorded
+  hours, so a budget is met to within one clip and `report()` says what was
+  actually delivered.
+- Resamples to one rate and returns one record shape whatever the source:
+  `audio, text, duration, language, split, source, clip_id, speaker_id`.
+
+## What it refuses to do
+
+- Mix splits. A training recipe reads train splits only, and asking it for a
+  test split raises rather than returning the audio.
+- Read a source whose recorded terms do not allow the recipe's purpose. The
+  permission answers travel inside the recipe and are checked again here, at
+  the point the data is fetched.
+- Keep audio. Clips are decoded, handed over and dropped.
+
+## What is not here yet
+
+- Per clip selection: length, speaker, topic and an exact hour budget all
+  need the clip index, which is the next step.
+- Adapters for zip archives, GitHub releases, cloud buckets, Dataverse and
+  Zenodo, and for local copies of datasets that need a signed agreement.
+- Writing shards or a NeMo manifest instead of streaming.
+
+A dataset this loader cannot yet fetch correctly is never written into a
+script. The builder page counts its hours, names it under the script, and
+says why it is not in the code.
+
+## Which datasets it can fetch
+
+The loader fetches whatever a recipe names, and a recipe is written by the
+builder page, which only writes a source it has checked: it asks each
+publisher's repository what configs and splits it really has and which
+columns a row carries, and records the date it asked. Nothing there is
+inferred from a naming convention, and a dataset whose train and test split
+is recorded somewhere other than the data itself is left out with that as
+the stated reason, rather than loaded on a guess.

afrospeech_recipes-0.1.0/afrospeech_recipes.egg-info/SOURCES.txt

@@ -0,0 +1,14 @@
+LICENSE
+README.md
+pyproject.toml
+afrospeech_recipes/__init__.py
+afrospeech_recipes/recipe.py
+afrospeech_recipes.egg-info/PKG-INFO
+afrospeech_recipes.egg-info/SOURCES.txt
+afrospeech_recipes.egg-info/dependency_links.txt
+afrospeech_recipes.egg-info/requires.txt
+afrospeech_recipes.egg-info/top_level.txt
+afrospeech_recipes/sources/__init__.py
+afrospeech_recipes/sources/audio.py
+afrospeech_recipes/sources/huggingface.py
+afrospeech_recipes/sources/tarball.py

afrospeech_recipes-0.1.0/afrospeech_recipes.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

afrospeech_recipes-0.1.0/afrospeech_recipes.egg-info/requires.txt

@@ -0,0 +1,7 @@
+requests>=2.28
+numpy>=1.23
+soundfile>=0.12
+scipy>=1.10
+
+[huggingface]
+datasets>=2.18

afrospeech_recipes-0.1.0/afrospeech_recipes.egg-info/top_level.txt

@@ -0,0 +1 @@
+afrospeech_recipes

afrospeech_recipes-0.1.0/pyproject.toml

@@ -0,0 +1,54 @@
+# The recipe loader, as published. The name, the licence and the metadata
+# below are the owner's decisions, recorded in
+# docs/decisions/2026-10-28-publishing-the-loader.md.
+#
+#     pip install "afrospeech-recipes[huggingface]"
+#     pip install -e "loader/[huggingface]"      # from a clone of this repo
+#
+# `datasets` is an extra rather than a dependency so the recipe rules and the
+# tar adapter can be tested, and a tar-only recipe run, without it.
+#
+# Versions are pinned to ranges, not exact releases: a free session installs
+# this beside whatever the host already has, and an exact pin would fight
+# Kaggle's and Colab's own numpy.
+
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "afrospeech-recipes"
+version = "0.1.0"
+description = "Load the speech data a recipe describes, from wherever it lives."
+readme = "README.md"
+requires-python = ">=3.10"
+license = "Apache-2.0"
+license-files = ["LICENSE"]
+authors = [{ name = "AfriModal AI", email = "kaykola.olaleye@gmail.com" }]
+keywords = ["speech", "asr", "african-languages", "datasets", "streaming"]
+classifiers = [
+  "Development Status :: 3 - Alpha",
+  "Intended Audience :: Science/Research",
+  "Programming Language :: Python :: 3",
+  "Topic :: Multimedia :: Sound/Audio :: Speech",
+  "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = [
+  "requests>=2.28",
+  "numpy>=1.23",
+  "soundfile>=0.12",
+  "scipy>=1.10",
+]
+
+# The one address that is public and will stay public. The source repository
+# is private, so a Source or Issues link would hand visitors a 404; both can
+# be added in a later release once there is something behind them.
+[project.urls]
+Homepage = "https://afrimodal-ai.github.io/"
+
+[project.optional-dependencies]
+huggingface = ["datasets>=2.18"]
+
+[tool.setuptools.packages.find]
+where = ["."]
+include = ["afrospeech_recipes*"]

afrospeech_recipes-0.1.0/setup.cfg

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