stronganchor/zazaki-genc-speech

Zazaki Genç Speech

This dataset is dedicated to the public domain under CC0 1.0 Universal. The speaker authorized public sharing without identifying attribution.

Publisher: zazacaogren.com.

This is a curated collection of short Genç Zazaki teaching recordings from one consenting speaker, with paired Latin orthography and IPA labels. It includes isolated words, introductions, and questions. Public records use opaque utterance, audio, and prompt IDs and a neutral speaker label. Original speaker names, account identifiers, recording titles, source URLs, and local paths are not included.

Versions and intended use

Configuration	Split	Examples	Purpose
historical_20260615	train	2,486	Original June 15 fine-tuning examples and labels
historical_20260615	validation	599	Original June 15 model-selection evaluation examples and labels
reviewed_20260719	full	4,694	Newer July 19 reviewed corpus with that snapshot's labels

The historical configuration preserves the original example membership, label strings, and training/validation assignment used by the June 15 models. Its 3,085 examples total approximately 93 minutes. The original eval split is named validation here because it was used for model/checkpoint selection. It is not an independently reserved final test set.

The newer configuration contains approximately 150 minutes. It is a reviewed source export, not the training set of those historical checkpoints, and it has no official training/test split. The configurations share 3,081 utterance IDs; 59 of those utterances have changed text or IPA labels. Do not concatenate the configurations as independent examples, or use the newer collection to evaluate a model trained on the historical collection without filtering overlap. Stable sample_id, audio_id, and prompt_id fields support overlap analysis. Exact counts, durations, and cross-version overlap statistics are in release_stats.json.

All examples come from one speaker. Scores on this corpus do not establish accuracy for unseen speakers, all Zazaki dialects, conversational audio, or long recordings. A new benchmark should hold out new recordings before model selection and group related prompts together.

Loading

Hugging Face repository: stronganchor/zazaki-genc-speech. Load the dataset with:

from datasets import load_dataset

historical = load_dataset("stronganchor/zazaki-genc-speech", "historical_20260615")
training = historical["train"]
validation = historical["validation"]

newer = load_dataset("stronganchor/zazaki-genc-speech", "reviewed_20260719", split="full")
example = training[0]
print(example["text"], example["ipa"])
audio = example["audio"]

The standard Parquet files contain embedded FLAC audio bytes. No custom dataset loader, source-site login, or external audio URL is required. Hugging Face Datasets exposes audio as its standard Audio feature; decoding dependencies vary with the installed Datasets version.

For local use, pass the matching files to the standard parquet loader. For example:

historical = load_dataset("parquet", data_files={
    "train": "data/historical_20260615/train.parquet",
    "validation": "data/historical_20260615/validation.parquet",
})

Columns

Field	Meaning
audio	Embedded FLAC, 16 kHz, mono, 16-bit PCM
sample_id	Opaque utterance identifier, stable across snapshots
audio_id	Opaque ID for identical decoded, normalized audio; exposes duplicate audio
speaker_id	Neutral speaker label; one value in this release
prompt_id	Opaque prompt-family identifier; related recording types share a prompt
source_id	Neutral collection identifier
text	Zazaki orthographic target, not an English translation
ipa	IPA transcription using this project's annotation conventions
recording_type	isolation, introduction, or question
duration_seconds	Duration of the released audio
review_status	Review status in the source snapshot
selection_method	Historical high-confidence filter or newer reviewed export
annotation_assistance	Specific assisted workflow where recorded, otherwise not_recorded_per_row

Labels retain the exact snapshot strings, including capitalization, punctuation, and Unicode normalization. text and ipa are separate targets; model comparisons should state their own normalization rules. Orthography reflects project conventions, including the historical labialized consonant spelling changes, and is not a claim of a universal Zazaki writing standard.

Annotation and audio processing

The historical set selected rows marked reviewed, without an open needs-review flag, and with every interlinear token's confidence at least 0.84. Nineteen additional questionable rows were removed before that historical training run. The newer snapshot uses the reviewed/no-needs-review source-export criterion and is restricted to the same verified speaker.

Transcription and review used a mixture of human review, model witnesses, linguistic/interlinear evidence, and some image-caption-assisted text checks. A reviewed flag is a workflow state; it does not mean that every label received independent double annotation. Some per-row assistance is recorded explicitly. not_recorded_per_row does not mean that automated assistance was absent.

Audio has been decoded, converted to mono at 16 kHz with FFmpeg, encoded as lossless FLAC containing 16-bit PCM, and checked by decoding every converted file. Existing MP3 sources remain limited by their original lossy encoding. No denoising, voice transformation, or silence trimming was applied for this release. The conversion removes source-file metadata. The released audio is a standardized representation, not a byte-for-byte copy of the original training files.

Historical duplicate examples are retained to reproduce the model's original example membership. The historical training split has one duplicate-audio pair with matching labels; its validation split has no duplicate audio. The newer corpus has five duplicate-audio groups containing 12 examples, including one group with differing labels. These source examples are preserved for transparency; resolve duplicate-label disagreements before treating the newer corpus as a clean training benchmark. Duplicate audio can be identified using audio_id; avoid splitting the same audio or related prompts across a new train/test boundary. The newer corpus may also include revised labels for the same utterance. Neither configuration is automatically deduplicated or a new validated benchmark.

Consent and identity

The speaker authorized public sharing without identifying attribution. This release removes direct identifying metadata and uses neutral IDs. The voice is not transformed; removing metadata does not make a human voice unrecognizable. As a nonbinding community request, please do not attempt to identify the speaker or link opaque IDs to private source records. This request does not add a legal restriction to CC0.

The release contains only the authorized speaker's teaching clips. No other speakers' corpora, third-party video recordings, private story collections, site snapshots, operational logs, or original export metadata are included.

License and citation

To the extent possible under law, the dataset contributors have waived copyright and related rights in this dataset under CC0 1.0 Universal. See LICENSE for the full legal text. Attribution is not required. This dataset dedication does not change the licenses of models trained with it.

If useful for your work, you may optionally cite the project:

zazacaogren.com. (2026). Zazaki Genç Speech. https://huggingface.co/datasets/stronganchor/zazaki-genc-speech

Project citation is appreciated as a nonbinding community request, not a condition of reuse. No identifying speaker credit is requested.