TriLingDATA/si-ta-en-asr

Sinhala-Tamil-English ASR Dataset

A combined speech recognition dataset for Sinhala, Tamil and English, built for fine-tuning multilingual ASR models such as Whisper. All audio is 16 kHz mono 16-bit WAV, transcripts keep cleaned punctuation and casing, and test/validation sets are held out by speaker.

A language can have several sources. Every row says where it comes from in the source column, so you choose which sources to train and evaluate on (see Select sources).

Splits

Split	Description	Sinhala (h)	Tamil (h)	English (h)	Total (h)	Clips	Speakers
full	All training data	166.8	222.8	250.8	640.4	352,026	~2,500
test_en	English test (LibriTTS-R + Svarah)	0.0	0.0	9.4	9.4	5,170	50
test_si	Sinhala test (OpenSLR + YouTube + BizBrains)	22.0	0.0	0.0	22.0	18,029	477
test_ta	Tamil test (IndicTTS + Common Voice)	0.0	28.8	0.0	28.8	13,723	636
val_en	English validation (LibriTTS-R + Svarah)	0.0	0.0	3.9	3.9	2,263	24
val_si	Sinhala validation (OpenSLR + YouTube + BizBrains)	5.8	0.0	0.0	5.8	4,440	104
val_ta	Tamil validation (IndicTTS + Common Voice)	0.0	22.3	0.0	22.3	12,699	50

• full is the training split. Shards are grouped by language and source, so shuffle before training.
• Test and validation sets never share clips or sentences with full.
• test_en, test_si, test_ta, val_en, val_si, val_ta contain more than one source; filter on source to evaluate one of them.
• To compare with Sinhala results reported before YouTube and BizBrains were added, evaluate test_si / val_si on source == "openslr_sinhala" only.
• Speaker counts do not include YouTube and BizBrains clips (their source has no speaker ids).

Hours (clips) per source

Split	openslr_sinhala	indic_tts_tamil	libritts_r	common_voice_tamil	youtube_sinhala	bizbrains_sinhala	svarah_english
full	161.7 h (133,534)	139.4 h (22,403)	241.8 h (142,252)	83.6 h (46,337)	4.4 h (2,453)	2.1 h (780)	7.3 h (4,267)
test_en	-	-	8.5 h (4,614)	-	-	-	1.0 h (556)
test_si	21.5 h (17,755)	-	-	-	0.3 h (176)	0.2 h (98)	-
test_ta	-	9.5 h (1,534)	-	19.3 h (12,189)	-	-	-
val_en	-	-	2.9 h (1,748)	-	-	-	1.0 h (515)
val_si	4.7 h (3,909)	-	-	-	0.8 h (433)	0.3 h (98)	-
val_ta	-	3.0 h (555)	-	19.3 h (12,144)	-	-	-

Columns

Column	Type	Description
id	string	Unique clip id: <lang>_<original id> (e.g. si_3bb010bb4e, ta_aea93c20_3434, ta_common_voice_ta_19916760); YouTube/BizBrains: si_yt_<hash> / si_biz_<hash> (first 10 hex characters of the SHA-1 of the audio); Svarah: en_svarah_<hash>
audio	Audio (16 kHz)	Speech recording, mono, 16-bit WAV
text	string	Transcript (Unicode NFC, punctuation kept)
language	string	Language of the clip: si, ta or en
duration	float32	Length in seconds
speaker_id	string	Speaker id (how it is formed for each source: see Sources)
source	string	Original dataset: openslr_sinhala, indic_tts_tamil, libritts_r, common_voice_tamil, youtube_sinhala, bizbrains_sinhala, svarah_english
in_split1	bool	True if the clip belongs to training subset split1 (only in full; always False in test/validation and for data added later)
in_split2	bool	True if the clip belongs to training subset split2 (only in full; always False in test/validation and for data added later)

Example row:

{
  "id": "si_3bb010bb4e",
  "audio": {"array": [...], "sampling_rate": 16000},
  "text": "ව්‍යාපාර කටයුතුවලින් විශේෂ වාසි ගෙන දෙයි.",
  "language": "si",
  "duration": 4.1,
  "speaker_id": "937dc",
  "source": "openslr_sinhala",
  "in_split1": False,
  "in_split2": False,
}

Select sources

Available sources:

Language	source	Shard files	Training hours	Kind of speech
Sinhala	openslr_sinhala	si-*	161.7	read sentences
Sinhala	youtube_sinhala	si_yt-*	4.4	Sinhala speech scraped from YouTube
Sinhala	bizbrains_sinhala	si_biz-*	2.1	Sinhala speech collected by BizBrains
Tamil	indic_tts_tamil	ta-*	139.4	spontaneous talk-style speech
Tamil	common_voice_tamil	ta_cv-*	83.6	crowd-sourced read sentences (originally MP3)
English	libritts_r	en-*	241.8	read audiobooks
English	svarah_english	en_svarah-*	7.3	Indian-accented English: read and spontaneous speech (speakers of 19 Indian languages)

from datasets import load_dataset

PICK = ["openslr_sinhala", "indic_tts_tamil", "libritts_r"]   # original sources
# PICK = ["openslr_sinhala", "common_voice_tamil", "libritts_r"]   # Common Voice for Tamil
# PICK = ["openslr_sinhala", "youtube_sinhala", "bizbrains_sinhala", "common_voice_tamil", "libritts_r"]   # all Sinhala sources
# PICK = ["openslr_sinhala", "indic_tts_tamil", "libritts_r", "svarah_english"]   # LibriTTS-R + Indian-accented English

full  = load_dataset("TriLingDATA/si-ta-en-asr", split="full")
train = full.filter(lambda s: s in PICK, input_columns="source")

test = load_dataset("TriLingDATA/si-ta-en-asr", split="test_si")
test = test.filter(lambda s: s in PICK, input_columns="source")

Note: language alone matches all sources of that language; use source to pick.

Download only the sources you need (saves disk)

Each shard file holds one source (<shard prefix>-NNNNN.parquet, see the table above). Loading by file pattern downloads only those files:

base = "hf://datasets/TriLingDATA/si-ta-en-asr/data/full"
prefixes = ["si", "si_yt", "si_biz", "ta_cv", "en", "en_svarah"]
train = load_dataset("parquet", data_files=[f"{base}/{p}-*.parquet" for p in prefixes], split="train")

Stream without downloading everything

ds = load_dataset("TriLingDATA/si-ta-en-asr", split="full", streaming=True)
ds = ds.filter(lambda s: s in PICK, input_columns="source")
print(next(iter(ds))["text"])

Filter by language or duration

sinhala_only = full.filter(lambda l: l == "si", input_columns="language")
short_clips  = full.filter(lambda d: d <= 15, input_columns="duration")

Decoding audio

With datasets>=4, audio decoding needs pip install torchcodec. Alternatively, read the raw WAV bytes:

import io, soundfile as sf
from datasets import Audio

ds = load_dataset("TriLingDATA/si-ta-en-asr", split="test_ta").cast_column("audio", Audio(decode=False))
array, sr = sf.read(io.BytesIO(ds[0]["audio"]["bytes"]))   # sr == 16000

Tips for fine-tuning Whisper

• Shuffle the training split: train = train.shuffle(seed=42).
• Set the language token per example from the language column (si / ta / en), not one language for the whole batch.
• Drop labels longer than 448 tokens. Sinhala and Tamil use many tokens per word, and some Tamil clips are close to 30 s:

  train = train.filter(lambda t: len(tokenizer(t).input_ids) <= 448, input_columns="text")
  

• Evaluation: transcripts contain punctuation and casing. For WER/CER, normalize both the prediction and the reference (lowercase, remove punctuation). Report CER as well as WER for Sinhala and Tamil, and report each source separately.

Training subsets

Smaller training sets of the original data are marked with boolean columns inside full. Subsets are sampled by speaker and may overlap each other (33,409 clips are in every subset). They contain only these sources: openslr_sinhala, indic_tts_tamil, libritts_r (only the original LibriTTS-R clips in en-00000…; the clips added later in en-00100… are not in any subset). common_voice_tamil, youtube_sinhala, bizbrains_sinhala and svarah_english are not in any subset.

Column	Sinhala (h)	Tamil (h)	English (h)	Total (h)	Clips
in_split1	57.2	47.5	28.5	133.3	71,504
in_split2	76.3	57.2	38.1	171.5	94,390

Sources

Language	Source dataset	source	Shard files	What was used	Speaker id
Sinhala	SPEAK-ASR/openslr-sinhala-asr	openslr_sinhala	si-*	All splits. Original test split kept as test_si; ~5 h of validation speakers as val_si, the remaining validation speakers added to training	source speaker_id
Tamil	psk/indic-tts-966h (tamil)	indic_tts_tamil	ta-*	All data; test and validation held out by speaker	file-name prefix (proxy; the source has no speaker column)
English	mythicinfinity/libritts_r (clean)	libritts_r	en-*	All of train.clean.360 and train.clean.100 (the first ~70 h of train.clean.360 in shards en-00000…, the rest added later in shards en-00100…); dev.clean (subsampled by speaker) and test.clean	source speaker_id
Tamil	TriLingDATA/Tamil_STT (Common Voice 22.0, Tamil)	common_voice_tamil	ta_cv-*	Source splits used as they are: train -> full, validation -> val_ta, test -> test_ta	cv_ + first 16 characters of client_id
Sinhala	SinhaSpeech/sinhala-asr-data (stratified_v6, source_dataset = youtube)	youtube_sinhala	si_yt-*	Source splits (speaker-disjoint) used as they are: train -> full, validation -> val_si, test -> test_si	none in the source: yt_unknown
Sinhala	SinhaSpeech/sinhala-asr-data (stratified_v6, source_dataset = bizbrains)	bizbrains_sinhala	si_biz-*	Source splits (speaker-disjoint) used as they are: train -> full, validation -> val_si, test -> test_si	none in the source: biz_unknown
English	ai4bharat/Svarah	svarah_english	en_svarah-*	The single test split, divided here by speaker (seed 42): ~10 % of hours -> test_en, ~10 % -> val_en, the rest -> full	svarah_ + hash of the speaker metadata (gender, age-group, primary_language, native_place_state, native_place_district, highest_qualification, job_category, occupation_domain; proxy, the source has no speaker id)

The full list is also in sources.json.

Processing

• Audio: resampled to 16 kHz, mono, 16-bit PCM WAV.
• Text: Unicode NFC; punctuation kept but cleaned (CSV double quotes, curly quotes, repeated punctuation and stray symbols fixed); Sinhala zero-width joiner kept only inside conjuncts (e.g. ්‍ය, ්‍ර); stray zero-width characters removed. The same cleaning is used for every source.
• Language labels: English-only clips found in the Sinhala/Tamil sources are labelled en.
• Filters: clips shorter than 1 s or longer than 30 s, empty transcripts and characters-per-second outliers were removed.
• Leakage control: test/validation sets are selected by speaker (sources added later: their own official splits); training clips whose sentence also appears in a test/validation set of the same language were removed. For YouTube/BizBrains, test/validation clips whose sentence already appears in the Sinhala training data were removed as well, and exact duplicate audio was dropped.
• Svarah split: Svarah is published as one test set. Here it is split by speaker as in the Sources table; training clips whose sentence appears in the Svarah test/validation part were removed.

Limitations

• Sinhala speaker overlap: 477 Sinhala speakers (26.2 h of test/validation audio) also appear in full, inherited from the source dataset's splits. Sinhala test scores may be slightly optimistic for unseen speakers.
• Speaker ids: IndicTTS has no speaker column; the file-name prefix is used as a proxy. YouTube and BizBrains have no speaker ids at all (yt_unknown, biz_unknown); their speaker-disjoint split comes from the source and cannot be re-checked here.
• Code-switching: IndicTTS Tamil transcripts contain English words in Latin script; Sinhala transcripts may contain some English words.
• Domains differ: OpenSLR (Sinhala) is read sentences; YouTube and BizBrains (Sinhala) are collected speech from other domains; IndicTTS (Tamil) is spontaneous talk-style speech; LibriTTS-R (English) is read audiobooks; Common Voice (Tamil) is crowd-sourced read sentences (originally MP3); Svarah (English) is Indian-accented read and spontaneous speech.
• Svarah is a benchmark: published Svarah results use all 9.6 h as test data; scores on this repo's Svarah part of test_en are not comparable to them, and models trained on full have seen other Svarah speakers.
• Clips longer than 30 s were removed, so very long utterances are not represented.

License

This dataset combines sources with different licenses. Use and redistribution must follow the license of each source dataset; check each source's dataset card before redistribution or commercial use. SinhaSpeech/sinhala-asr-data (YouTube, BizBrains) does not state a license, and YouTube audio may be subject to the rights of the original uploaders. ai4bharat/Svarah is released under CC BY 4.0: credit AI4Bharat and cite the Svarah paper.

Citation

If you use this dataset, please cite the source datasets listed above.

Svarah: T. Javed et al., Svarah: Evaluating English ASR Systems on Indian Accents, INTERSPEECH 2023 (arXiv:2305.15760).