Reza2kn/shenava-koochik-number-stress-36k
Shenava Koochik Number Stress 36K This repository contains the 36,000-row Persian number-stress text curriculum and an audio-backed subset of 6,669 short clips generated with Gemini TTS. The Dataset Viewer default configuration is the 6,669-clip audio subset. It exposes audio, text, category, and duration_s; the metadata links each row to its WAV using a relative file_name such as audio/num-000000.wav. The complete text-only source remains available at data/train.jsonl (36,000… See the full description on the dataset page: https://huggingface.co/datasets/Reza2kn/shenava-koochik-number-stress-36k.
Shenava Koochik Number Stress 36K
This repository contains the 36,000-row Persian number-stress text curriculum and an audio-backed subset of 6,669 short clips generated with Gemini TTS.
The Dataset Viewer default configuration is the 6,669-clip audio subset. It exposes audio, text, category, and duration_s; the metadata links each row to its WAV using a relative file_name such as audio/num-000000.wav.
The complete text-only source remains available at data/train.jsonl (36,000 rows, with text and category fields).
The text utterances are constrained to 2–8 whitespace-separated words and contain no Arabic/Persian digits or Latin letters. Categories cover clock time, dates, cardinals, ordinals, arithmetic, money, percentages, fractions, ranges, measures, and sequence IDs.
The text contexts were requested from z-ai/glm-5.3-flash using short number-centric in-context examples. Rows that failed hard constraints were completed with deterministic short templates and retained in the same category distribution.