Scicom-intl/Synthetic-User-Turn-TTS

Synthetic Malaysian Telco Call-Centre Speech

Text and speech for training ASR on Malaysian telco call-centre calls.

The dialogues are written by an LLM in the style of real Unifi call-centre recordings. They contain no real customer data. The audio is made with voice conversion. A reference voice reads each line, Whisper transcribes the result, and a clip is published only if the transcript matches the text.

There are 262 hours of audio in 97,371 clips across 14 audio subsets, plus 13 text subsets.

from datasets import load_dataset

text  = load_dataset("Scicom-intl/Synthetic-User-Turn-TTS", "dialogues_v2", split="train")
audio = load_dataset("Scicom-intl/Synthetic-User-Turn-TTS", "spelling_texts", split="train")

Subsets

Audio

subset	clips	hours	gate	content
hard_entities_mandarin_texts	23,576	62.9	WER ≤ 0.10, CER ≤ 0.05, digits exact	Malaysian Mandarin turns full of IDs, phones and amounts
dialogue_calls	6,262	47.2	per turn: WER ≤ 0.10, CER ≤ 0.05, digits exact	whole calls, two voices, 30 s window
spelling_texts	19,116	30.6	spelled content exact	callers spelling names, emails and numbers
spelling_texts_filter	15,706	23.5	as spelling_texts, and text_written carries every spoken word	spelling_texts for training on text_written
hard_entities_tamil_v2_texts	9,573	23.2	WER ≤ 0.50, CER ≤ 0.15, digits exact	Malaysian Tamil, second set
hard_entities_v2_texts	8,346	22.6	WER ≤ 0.10, CER ≤ 0.05, digits exact	entity turns, 64% Mandarin
synthetic	5,002 + 200 test	14.4	WER ≤ 0.05, CER ≤ 0.03	customer turns, DNSMOS voices
synthetic_podcast	8,458 + 200 test	13.6	WER ≤ 0.05, CER ≤ 0.03	customer turns, podcast voices
hard_entities_tamil_texts	4,992	12.0	WER ≤ 0.50, CER ≤ 0.15, digits exact	Malaysian Tamil
hard_entities_calls	1,798	11.9	per turn: WER ≤ 0.10, CER ≤ 0.05, digits exact	whole entity calls
hard_entities_texts	4,950	10.7	WER ≤ 0.10, CER ≤ 0.05, digits exact	entity turns, 58% Malay
hard_entities_tamil_calls	979	6.4	per turn: WER ≤ 0.50, CER ≤ 0.15, digits exact	2-turn Tamil calls
revolab_style	3,167	5.1	WER ≤ 0.05, CER ≤ 0.03	11 speech registers (news, drama, podcast...)
clean	911	1.5	WER = CER = 0, speaker similarity, UTMOS, speaking rate	both roles, strictest quality
revolab_style_si	241	0.3	WER ≤ 0.12, CER ≤ 0.06	street-interview register

All audio is 24 kHz mono. Every row keeps its own scores, so you can always filter to a stricter set.

Text

subset	dialogues	messages	content
default	4,260		source dialogues, 213 topics
dialogues_v2	7,780		source dialogues, 389 topics
dialogues_normalized	12,040		the two above merged, plus a spoken form
hard_entities	670	5,613	entity-dense calls, 60% ms / 35% en / 5% zh
hard_entities_v2	1,300	10,179	entity-dense calls, 65% zh
hard_entities_tamil	1,300	9,549	entity-dense calls, 100% Tamil
hard_entities_tamil_v2	2,599	19,294	entity-dense calls, 100% Tamil
hard_entities_mandarin	3,900	31,044	entity-dense calls, 100% Mandarin
hard_entities*_normalized	same	same	each set above, plus a spoken form

How a clip is made

target text + reference voice
        |
        v
voice conversion  ->  Whisper  ->  compare with target  ->  pass: publish
        ^                                                    |
        +------------ fail: new take, new temperature -------+

1. A voice-conversion model speaks the target text in the voice of a reference recording. It needs the reference audio and its transcript.
2. Whisper transcribes the result.
3. The transcript is scored against the target. A take that fails is regenerated at a different temperature. API errors do not use up the take budget.
4. Only clips that pass are published. Texts that never pass are left out.

Scoring rules:

• Case, punctuation and filler spellings (erm, um, uh) are normalized first.
• Chinese is scored per character. Without spaces, a whole sentence would be one word.
• Digit runs are masked out of WER and CER, then checked exactly on their own (digits_ok). Whisper groups digits freely, so 012-1579129 and 0121579129 would otherwise count as different words. A single wrong digit, on the other hand, barely moves CER.
• On a failing take, Whisper's output is rewritten into spoken form by the same LLM prompt that wrote the target, then scored again. RM1,250.50 and seribu dua ratus lima puluh ringgit lima puluh sen share no digits, so masking alone cannot match them.

Reference voices come from Scicom-intl/Multilingual-TTS-DNSMOS (84 source corpora, DNSMOS OVRL ≥ 4.0) and Scicom-intl/Clean-Podcast (Malaysian and Singaporean podcasts). Most voices are not Malaysian. Cross-lingual voice conversion works well here: Chinese targets passed more often than English ones.

synthetic and synthetic_podcast

Single customer turns. Both share one schema:

column	type	description
audio	audio	24 kHz mono wav
text	string	the target text
transcription	string	what Whisper heard
wer, cer	float	transcription against text
duration	float	seconds
takes	int	takes spent before this one passed
temperature	float	sampling temperature of the accepted take
speaker, speaker_config	string	reference voice and its source corpus
speaker_ovrl	float	DNSMOS OVRL of the reference (null for podcast voices)
reference_text	string	transcript of the reference
reference_duration	float	reference length in seconds
sample_rate	int	24000

synthetic: text from Scicom-intl/Malaysian-Call-Center-Language-Switching. Voices from Multilingual-TTS-DNSMOS, spread across source corpora so no corpus dominates. Every clip has its own speaker: 5,102 voices from 77 corpora. Mean WER 0.0108, mean CER 0.0060, and 63.4% of clips score zero on both. It took 3.78 takes per clip on average, up to 10.

synthetic_podcast: text from this dataset's default subset. Voices are 15-second Clean-Podcast chunks, with that dataset's STT output as the reference transcript. Speakers repeat here. Clean-Podcast has about 676 source videos, so each voice appears roughly ten times, each time from a different segment. Use synthetic if you need one speaker per clip.

Both have two extra splits of 100 clips. test is a held-out sample. test_livekit is the same 100 clips played through a real LiveKit call (48 kHz Opus over WebRTC, silero VAD), saved as the 16 kHz audio the STT plugin received. Rows line up with test by position, so you can compare clean and degraded audio directly.

clean

911 clips, 1.5 hours, both roles (460 customer, 451 agent), 471 English, 399 Malay and 41 Chinese. Text from default and dialogues_v2.

This is the strictest subset. A clip must pass four gates:

gate	threshold	why
transcript	WER ≤ 0.05, CER ≤ 0.03	every published clip is in fact at 0
speaker similarity	TitaNet cosine ≥ 0.60 against the reference	it must sound like the reference voice
naturalness	UTMOSv2 ≥ 3.0	it must sound like real speech
speaking rate	Latin 9.5-30, CJK 3.6-14 chars/s	catches clips that keep talking after the text ends

The rate gate exists because WER is blind to extra audio. A take can say the sentence and then babble, and Whisper simply does not transcribe the tail. Each text gets 20 takes, split across several reference voices. Similarity depends more on the voice than on the take, so trying more voices pays off more than rolling more takes.

Mean similarity is 0.650 and mean UTMOSv2 is 3.31. For scale: two clips of the same real podcast speaker score about 0.84, and two different speakers about 0.64. A cloned voice against its own reference usually lands at 0.50 to 0.65.

Audio is leveled to -20 dBFS active RMS with a -1 dBFS peak ceiling. This is gain only, no compression. gain_db records the gain. Extra columns: similarity, utmos, gain_db, role, script, speaker_utmos, and reference_audio (the reference clip itself).

revolab_style and revolab_style_si

Text written to match the 11 registers of Revolab/ASR-Benchmark-Public: fleurs, commonvoice, news, parliament, podcast, drama, street interview, animation, read speech, telephony and short inputs. 3,167 clips, 5.1 hours, 1,911 voices, mean WER 0.0034.

Entities (names, places, amounts) were taken from the benchmark's references. Sentences were not. Any line sharing a 6-word sequence or a whole clause with a reference was dropped. Only 2 of 4,414 candidates hit that check.

Training on this makes that benchmark in-domain. It is no longer a held-out test. In our runs it also did not lower WER on the benchmark, which suggests the remaining errors there are acoustic, not vocabulary.

revolab_style_si holds 241 street-interview clips at a looser WER ≤ 0.12. At WER ≤ 0.05 a 14-word line allows zero errors, and Whisper tends to drop fillers like aa and um, so only 10% of street-interview texts passed. The looser gate passes 60%, but some labels then contain fillers the audio may not have. In our runs these clips made street-interview WER worse. Use with care.

The hard_entities family

Built for one failure: numbers and names. A model trained on generic call text still mishears the content that decides a call, for example 102105 as 2105, 950 as 150, or Credit card as Kedipa. Those turns are exactly the ones a transcript gate throws away, so ordinary synthetic corpora rarely contain them.

Every dialogue carries an IC number, a mobile number, a case or reference number, and usually an RM amount, written into a normal conversation. Dialogues open on the caller's line. They cover the same 15 call-centre categories as default.

The five text sets (hard_entities, _v2, _tamil, _tamil_v2, _mandarin) are different dialogues, not versions of each other. Each round was written at a different language mix. Load all five for all 9,769 dialogues.

Schema matches default. The _normalized variants add normalized_messages, the spoken form fed to the voice model. messages stays the written form and is the ASR label. num_turns counts messages in the v2 sets and user turns in hard_entities, so prefer len(messages).

The *_texts audio subsets hold one clip per unique turn, deduplicated by (language, text). Both roles are voiced and each clip has its own reference voice.

column	type	description
audio	audio	24 kHz mono wav, unprocessed
text	string	written form, the ASR label
spoken_text	string	spoken form fed to the voice model
role	string	user or assistant
wer, cer	float	of the accepted take
digits_ok	bool	the digit sequence was transcribed exactly
voice, voice_config	string	reference voice and its source corpus
topic, category, language	string	from the source dialogue
source_dialogue_id	string	id of the source dialogue
attempt, duration, sample_rate		take index, seconds, 24000

hard_entities_texts: 4,950 clips, 10.7 hours. 2,849 Malay, 1,863 English, 238 Chinese.

hard_entities_v2_texts: 8,346 clips, 22.6 hours. 5,311 Chinese, 1,562 English, 1,473 Malay. Mean WER 0.0160, CER 0.0110. 738 of 9,084 source texts (8.1%) never passed and are missing. This set shares only 1 utterance with hard_entities_texts, so the two can be combined freely.

All clips in every *_texts subset have digits_ok = true.

hard_entities_mandarin_texts

23,576 clips and 62.9 hours of Malaysian Mandarin, the largest subset here. 10,912 customer turns and 12,664 agent turns, 8,227 voices, all 3,900 source dialogues covered, median clip 8.8 s.

It uses the standard WER ≤ 0.10 / CER ≤ 0.05 gate because Whisper handles Mandarin well. Before the run, 95% of sampled turns cleared it. In the final set, mean WER and CER are 0.0113, the median is 0.000, and 62% of clips have zero character errors.

On a mostly-Chinese turn WER and CER are the same number, since both are scored per character. They differ only on English-heavy turns, which is why max WER (0.10) is above max CER (0.05).

All 11,083 digit-bearing clips match their digits exactly. The spoken form reads digits in Chinese (016-4432109 becomes 零一六四四三二一零九). Whisper writes them back as digits, so the round trip is checked.

The Tamil subsets

15,544 clips and 41.2 hours from two separate Tamil dialogue sets.

subset	clips	hours	shape
hard_entities_tamil_texts	4,992	12.0	one clip per turn
hard_entities_tamil_v2_texts	9,573	23.2	one clip per turn
hard_entities_tamil_calls	979	6.4	2-turn call, as in dialogue_calls

The *_texts subsets have 7,539 customer and 7,026 agent turns, 8,305 voices, median clip 7.8 s.

They use WER ≤ 0.50 / CER ≤ 0.15, not 0.10 / 0.05. The audio is not worse. Whisper is bad at Malaysian Tamil, and the gate has to match the instrument. We ran real human Tamil speech from Scicom-intl/Malaysian-Tamil-Emilia through the same Whisper. It scored more than twice as badly as our synthetic clips. Even the easiest tenth of the human recordings only reach CER 0.130, so a CER 0.05 gate would reject almost all real Tamil speech too.

The remaining WER is mostly spelling and word boundaries (பெயர் vs பேர், சரிபார்க்கிறேன் vs சரி பார்க்கிறேன்). Tamil joins words together, so one boundary choice costs two word errors but barely moves CER. Filter on cer and digits_ok, not on wer.

Digits are not relaxed. All 1,994 digit-bearing clips in _texts, all 3,813 in _v2_texts, and all 634 digit-bearing turns in _calls match exactly.

Each *_texts subset was built in two passes. The first keeps about half the turns. The second retries only the rejects with a different voice and recovers about a quarter of them. Tamil accuracy depends more on which voice reads the line than on how many times it is retried.

spelling_texts

Callers spelling things out: names, family names, emails, phone and account numbers, car plates, user IDs and reference codes. In production the agent asks "can you spell that?" and the STT loses the answer. H, U, S, E, I, N, Husein comes back as Hussein Hussein, and triple zero five never becomes 0005.

19,116 clips, 30.6 hours, median 5.2 s, 8,034 voices. 10,385 Malay, 6,850 English, 1,881 Chinese.

Each text is built from parts, so every field is exact by construction:

column	example
spoken_text (fed to the voice model)	Emel saya H, U, S, E, I, N, double kosong, tujuh, at gmail dot com.
text (verbatim label)	Emel saya H U S E I N 0 0 7 at gmail dot com.
text_written (written label)	Emel saya husein007@gmail.com.
spelled (what the gate checks)	HUSEIN007

Pick a label. text keeps the spelled letters as spaced capitals. text_written collapses them into the address or number as you would type it. Which one to train on is your choice. In text, "double S" is written S S, and real words such as at, dot and underscore stay words.

What is in it:

• kind: email 6,713, name 5,520, phone or account number 2,401, code 2,377, short code 815, family name 761, surname 529.
• subkind: name origin (Malay 6,108, Chinese 3,318, Tamil 2,035, international 1,404, mixed 658) or number type (mobile 1,561, account 840, plate 811, reference 801, user ID 765).
• spell_sep: how letters are separated. comma is H, U, S (7,723). space is H U S, E I N (5,699). chunk is groups of 3-4, S H A, F I Q (5,694).
• style for names: letters (spelling only), letters_word (spelling then the name), word_letters (the name then spelling), phonetic (H for hotel).
• About a third of the texts use double or triple for repeated characters.

The gate is different from the other subsets. WER over spelled letters mostly measures Whisper's formatting. The same audio comes back as H-U-S-E-I-N, HUSEIN or H, U, S, and the letter U sometimes comes back as you. So a clip passes on two checks:

1. The spelled content must match exactly. Every spelled character is read out of both the target and the transcript, whatever the format. Digit words in English, Malay and Chinese are folded, double and triple are expanded, and O equals 0 because "oh" means both. One wrong or missing character fails the clip.
2. The rest of the sentence may differ by at most 2 tokens, or 25% of its tokens if that is more. These carrier phrases are only 0 to 6 words long, so a rate gate would fail a clip for one respelled name. carrier_errs gives the count. 12,636 rows have zero. Most of the others are spelling conventions, like Okay vs OK or emel vs email. Filter on carrier_errs == 0 if you need the whole sentence exact.

Each take is judged by Whisper twice. Unprompted, Whisper often merges the spelled letters into the word: My name is H, U, S, E, I, N, Husein comes back as My name is Hussein Hussein. The first pass gives Whisper a prompt showing a made-up name and email spelled out letter by letter, so it writes letters too. The prompt never contains the target, so the content still has to come from the audio. The second pass uses no prompt. 17,136 clips passed on the prompted pass and 1,980 on the plain one. judge says which, and hyp is that transcript.

The first pass kept 89.2% of 20,000 texts. Two more passes retried only the rejects with new voices and recovered 945 and then 329 more, for 95.6% in total. The weakest cell is Malay phone and account numbers.

spelling_texts_filter

spelling_texts without the rows whose text_written leaves out words the voice says. Train on audio + text_written. On those rows a model would learn to delete speech.

15,706 of 19,116 clips, 23.5 hours. 8,573 Malay, 5,618 English, 1,515 Chinese. Built from spelling_texts at revision d0717e9d; every column is unchanged.

removed	rows	text	text_written
phonetic alphabet (style == phonetic)	698	Okay, N for nancy A for apple D for dog I for ice A for apple.	Okay, NADIA.
then, and then or 然后 between spelled names	1,653	Boleh, Devi, D E V I, then Nagarajan, N A G A R A J A N.	Boleh, Devi, DEVI, Nagarajan, NAGARAJAN.
all small letters, semua huruf kecil or 全部小写 after an email	1,059	P R I Y A 8 8 at yahoo dot com, all small letters.	priya88@yahoo.com.

The rule. Reduce text and text_written to their letters, digits and CJK characters, in order. In text, drop the spoken symbol words (at, dot/titik, underscore, dash) only when text_written holds that symbol. Keep the row when the two strings are equal. Phonetic rows always go. Every rejected row falls into one of the three groups above.

Checked, nothing found. In every row the spoken at, dot, underscore and dash match the symbols inside the written address one for one, and the digits match. spoken_text says double L where text writes L L: that is how people spell aloud and the label is right, so those rows stay. The spelled column writes the letter O as 0 and joins multi-name spellings; it is the gate's key, not a label.

Multi-turn calls

dialogue_calls, hard_entities_calls and hard_entities_tamil_calls put a whole stretch of a call into one clip: agent and customer turns, two different voices, and silence between them.

subset	clips	hours	median	languages
dialogue_calls	6,262	47.2	30.0 s	3,627 Malay, 2,236 English, 399 Chinese
hard_entities_calls	1,798	11.9	23.9 s	964 Malay, 749 English, 85 Chinese
hard_entities_tamil_calls	979	6.4	23.8 s	Tamil

Why: a model finetuned only on tightly cut single turns makes up text when it hears long silence. Real call recordings are fixed windows with sparse speech. On a real call-centre test set the median clip is 30 s, 26% of clips are under half speech, and 24% have a silence gap over 2 s.

• Every clip is at most 30 s, Whisper's input window.
• Silence is filled with room tone, not digital zeros. Each clip has a random late start, gaps between turns and a tail.
• The two voices always come from two different speakers in two different source corpora.
• Turns are chosen by content. A long dialogue is split into consecutive windows of whole exchanges, so every turn ends up in some clip and an IC readback late in the call is not cut off.
• Each turn is scored on its own, and every turn in a clip must pass. dialogue_calls clips hold 2 to 6 turns, hard_entities_calls 2 to 4.
• Audio is unprocessed: no resampling, no loudness normalization, no telephone band. Channel effects are left to train-time augmentation.

column	description
audio	24 kHz mono
text	written transcript of the turns in the clip, the ASR label
spoken_text	spoken form fed to the voice model
original_speaker_assistant, original_speaker_user	<config>/<speaker> of the two voices
assistant_config, user_config	source corpus of each voice
voice_distance	speaker-embedding distance between the two voices
turns	per-turn detail: role, text, WER/CER, duration, voice
turns_used, turns_total, window_start	which part of the dialogue this clip covers
n_number_turns, n_entity_turns	turns in the clip carrying numbers or entities
onset, gaps, tail, speech_ratio, duration	the silence layout
telephone_band, role_levels	always false: no band-limiting, no level difference between roles

hard_entities_tamil_calls holds only 2-turn calls. At about a 50% pass rate per Tamil turn, a 4-turn clip would pass about 6% of the time.

Source dialogues

default: 4,260 short call snippets, one row per dialogue.

• topic: the seeded call-centre topic
• category: one of 15 categories (billing, technical support, installation, account management, complaints and others)
• language: english, malay or chinese, the main language of the call. Malay and Chinese rows code-switch with English terms, as real Malaysian calls do.
• num_turns (1-3): agent and customer exchanges
• messages: alternating assistant (agent) and user (customer) turns, starting with the agent

213 topics, 20 dialogues each, split 7 English / 12 Malay / 1 Chinese to match the real call centre (about 35% / 60% / 5%). Names, numbers, dates and amounts are all made up.

dialogues_v2: same recipe, 389 topics (the 213 above plus 176 new ones, no overlap). id is a new UUID per row. Its customer turns give 14,472 unique utterances, against 7,667 in default.

dialogues_normalized: default and dialogues_v2 merged (12,040 rows, source_config says which), plus normalized_messages, every turn rewritten into the words a speaker would say. 880214-10-5566 becomes eight eight zero two one four dash one zero dash five five six six. messages stays the written form and is the ASR label.

Without this split, a raw 880214-10-5566 can be neither spoken by a TTS nor checked by an ASR round trip, so turns with IC, phone or account numbers were being dropped from speech corpora. Those are the turns an IVR most needs to get right.

Normalization is language-aware, because a bare number has no language and would otherwise come back in Chinese numerals inside a Malay call. 8,247 turns were rewritten. Every rewrite is checked for dropped words, leftover digits, large length changes and script drift. A rewrite that fails keeps the original text.

Caveats

• Passing the gate does not mean perfect audio. The gate limits how much Whisper disagrees with the text. It does not measure how natural the clip sounds. Listen to a sample before training.
• Some clips in the older subsets peak at full scale. Add headroom if that matters.
• Hard text is under-represented. Texts that never passed are not published, and those are mostly the ones with unusual names or long numbers. The source text subsets are complete. The audio subsets are not.
• The checker limits the corpus. Every clip was verified by Whisper, so the corpus cannot contain what Whisper cannot already hear. Names Whisper always misspells and fillers it always drops get filtered out.

Licensing and provenance

The dialogue text is synthetic and contains no real customer data. Reference voices come from the upstream corpora linked above and keep their original licences.