Hawrami Speech (hawrami_speech)

Studio-recorded Hawrami (Hewramî) read speech with sentence-level transcriptions: 4,977 utterances over ~4 hours of audio, with speaker_id labels covering 102 distinct speakers.

At a glance

Rows	4,977 — train 4,773 / test 204
Columns	audio, sentence, gender, language, original_full_path, duration, speaker_id
Parquet on disk	467.0 MB
Audio format	WAV (files like voice__24773.wav)
Language	Hawrami (language is hawrami in every row)
Task	automatic speech recognition

+---------+-------------+------------------+------------+------------+--------------+--------------+--------------+
| Split   |   Sentences |   Duration (hrs) |   Speakers |   Male (%) |   Female (%) |   Avg. Sent. |   Avg. Dur./ |
|         |             |                  |            |            |              |       Length |     Sent (s) |
+=========+=============+==================+============+============+==============+==============+==============+
| Train   |        4773 |             3.89 |         99 |       56.6 |         43.4 |         33.4 |         2.94 |
+---------+-------------+------------------+------------+------------+--------------+--------------+--------------+
| Test    |         204 |             0.16 |         19 |       68.4 |         31.6 |         31.8 |         2.78 |
+---------+-------------+------------------+------------+------------+--------------+--------------+--------------+

Field profile

Measured on the test split:

Column	Value
sentence	mean 35 chars, median 33, max 91; Arabic script
duration	0.98 – 6.33 s
gender	male / female
speaker_id	SPEAKER_100 style ids — 102 distinct across the dataset
original_full_path	36 chars, e.g. hawrami_speech/test/voice__24773.wav
language	always hawrami

Sample sentences:

ئەی دنیاء
ئەک هەی دماڎ بڕیوْ

Note the Hawrami-specific letters (ڎ, ۆ combos, and the ۊ/و variation) — a Sorani normaliser will rewrite some of these, so do not apply Sorani orthography rules blindly to this data.

Usage

from datasets import load_dataset

ds = load_dataset("razhan/hawrami_speech")
print(ds["train"][0]["sentence"], ds["train"][0]["speaker_id"])
print(ds["test"][0]["duration"])

Speaker counts (all 4,977 rows)

Counting speaker_id over the whole dataset gives 102 distinct speakers, not the 118 the 99 + 19 figures above suggest — so the train and test splits share speakers and the provided split is not speaker-disjoint. Gender over the same full file: male 2,701 (54.3%) / female 2,276 (45.7%).

If you need a speaker-disjoint evaluation, re-split on speaker_id yourself:

speakers = sorted(set(ds["train"]["speaker_id"]) | set(ds["test"]["speaker_id"]))
held_out = set(speakers[:12])
train = ds.filter(lambda r: r["speaker_id"] not in held_out)

How it was built

Recorded in studio conditions (read speech), then segmented per sentence. The original_full_path column keeps the recording path so clips can be traced back to their session.

Known limitations

• Read speech only, and studio-clean: no noise, phone channel, or spontaneous speech.
• Small (~4 hours) — fine-tuning scale.
• Hawrami has no single standard orthography; the written form here mixes ڎ/د and ۊ/و variants, so character error rates will vary with your normalisation.
• Sentence lengths are short (mean 33 chars), so long-form recognition is untested.
• No dialect/region labels. Speakers are roughly balanced overall (54.3% male / 45.7% female across all rows), but the provided test split skews male at 68.4%.
• The provided train/test split is not speaker-disjoint — see the speaker counts above — so per-split scores partly measure speaker memorisation.

Related

• razhan/DOLMA-speech (hawrami config) — 11,429 Hawrami utterances with English translations, a different recording effort.
• razhan/hac_to_en — Hawrami → English text pairs.
• razhan/horami — Hawrami article corpus.