fleurs-r-neucodec-all-languages

# FLEURS-R NeuCodec All Languages FLEURS-R metadata, source audio and precomputed NeuCodec speech tokens for 102 locales, plus a `speaker` label FLEURS itself does not ship. ## Layout - `data/{locale}-{split}.parquet` — metadata, one row per utterance (this is what the viewer shows). - `audio/{locale}-{split}.zip` — source FLEURS-R audio, 24kHz mono PCM16 WAV, members named `audio/{locale}/{split}/{id}.wav` (the `path` column). - `neucodec/{locale}-{split}-rank{N}.zip` — NeuCodec tokens, members named `neucodec/{locale}/{split}/{id}.pt` (the `neucodec_path` column), each a `{'source': str, 'codes': IntTensor[1, 1, T]}` torch file at 50 tokens/s. ## Columns | column | what | |---|---| | `id` | FLEURS utterance id | | `path` | audio member path inside `audio/{locale}-{split}.zip` | | `neucodec_path` | token member path inside the matching `neucodec/` zip | | `filename` | bare `{id}.wav` | | `sentence` | raw transcription | | `normalized_text` | lowercased, punctuation-stripped transcription | | `characters` | space-separated character transcription | | `locale`, `split` | e.g. `en_us`, `train` | | `gender` | FLEURS speaker gender label | | `speaker` | **derived** voice identity, `{locale}_spkNN` (see below) | | `num_samples` | sample count of the original 16kHz FLEURS audio | ## Speaker labels FLEURS and FLEURS-R ship `gender` but no speaker identity. `speaker` here is derived: every utterance is embedded with TitaNet-L ([titanet-vectors-fp16](https://github.com/Scicom-AI-Enterprise-Organization/titanet-vectors-fp16)), and the vectors are clustered per locale with agglomerative clustering (cosine distance, average linkage, threshold 0.4) **within each gender**, so a cluster never merges voices FLEURS already labels apart. Clusters smaller than 5 utterances are folded into their nearest centroid, and ids are numbered by descending cluster size, so `{locale}_spk00` is the most-recorded voice of that locale. They are clusters, not ground truth: treat them as consistent voice tags for conditioning, not as verified speaker identities. Across 102 locales / 279,496 utterances: 5–43 speakers per locale (median 11), mean silhouette 0.60. ## Loading ```python from datasets import load_dataset ds = load_dataset("malaysia-ai/fleurs-r-neucodec-all-languages") # all locales ds = load_dataset("malaysia-ai/fleurs-r-neucodec-all-languages", "en_us") # one locale ``` Audio and tokens are the zips above — `path` and `neucodec_path` name the member inside `audio/{locale}-{split}.zip` and `neucodec/{locale}-{split}-rank*.zip`. ## Source [google/fleurs-r](https://huggingface.co/datasets/google/fleurs-r) (FLEURS-R, CC BY 4.0), the restored variant of [google/fleurs](https://huggingface.co/datasets/google/fleurs).