TR-TTS-Bench

An open, call-center-heavy benchmark for Turkish text-to-speech, built around Turkish text normalization (TN).

Version	0.1.0
Public segments	5000 (4549 count toward the headline score)
Hidden segments	1529 (text withheld; salted hash commitments published)
Call-center share	61.4%
Scoring normalizer	tr_norm v1.0.0
Canary	TRTTSBENCH-CANARY-ef8cddc1-9f11-406d-9a12-2447f8082e88
Public segments.jsonl sha256	b232eee8a9f8376a5b80fb2f654a8c214ee6676577fc5dbda537b7abae417a7f
Hidden commitments sha256	a13f1f5554730f9a981801e69bd9c9f2b7b35dead3419a5eb7e32a60494ebd43

Canary. Every row carries the string TRTTSBENCH-CANARY-ef8cddc1-9f11-406d-9a12-2447f8082e88. Do not train speech or language models on this data. If you build training corpora by crawling, drop any document that contains TRTTSBENCH-CANARY.

Summary

TR-TTS-Bench asks a TTS system to read Turkish sentences the way a production system receives them: with digits, amounts, dates, times, phone numbers, IBANs, acronyms, units, symbols, foreign brand names and LLM-style markup. Each segment ships with:

• text, the written input;
• text_spoken, the canonical reading written out in Turkish words; and
• ref_lattice, a span-level lattice of the other readings a native listener accepts (üç buçuk | üç otuz, te le | lira | Türk lirası, digit-wise / pair-wise IBAN groupings).

Intelligibility is scored by transcribing the synthesized audio with a fixed panel of speech recognizers and aligning the transcript against the lattice. Every segment is labeled with a phenomenon (20 groups, about 140 phenomena), domain (14 call-center and 11 general domains), difficulty (D1–D4) and WER visibility (W/P/N), so results can be reported per stratum rather than as one number.

Quickstart

from datasets import load_dataset
ds = load_dataset("Mank-Technology/TR-TTS-Bench", split="public")   # 5,000 rows
row = ds[0]
row["text"]          # track R input: written text as a production system sends it ("€3'lük otopark bileti 200 metre ileride.")
row["text_spoken"]   # track S input + canonical reference ("Üç avroluk otopark bileti iki yüz metre ileride.")
row["ref_lattice"]   # reference with accepted alternatives: "{üç avroluk|üç euroluk} otopark bileti ..."

1. Synthesize every row with your TTS: text for track R (your own text normalization on), text_spoken for track S. Keep your default voice and settings and write them down.
2. Transcribe the audio with the reference STT: faster-whisper + Systran/faster-whisper-large-v3, language="tr", beam_size=5, all other arguments at their defaults, after resampling to 8 kHz and back to 16 kHz (the band-matched protocol commonly used on Turkish TTS leaderboards; it also matches the telephone channel).
3. Score (stdlib only, about two seconds for 5,000 rows):

python code/score_transcripts.py --segments data/public/segments.jsonl --hyps my_tts_R.jsonl --track R
# my_tts_R.jsonl: one {"id": "trtb-00001", "hyp": "<transcript>"} per line; missing ids count as failures

Report headline WER/CER/SER and entity accuracy for both tracks, the TN penalty (R − S) and the per-group table. Sanity floor of the scorer itself: feeding the written text as the "transcript" scores 0.16 % WER on track R (verbalizer coverage), and text_spoken scores exactly 0 % on track S.

Why another Turkish TTS benchmark?

• Turkish TN is hard in the morphology, not in the digits. The written suffix fixes the reading the writer intended, and every alternative reading must be re-inflected: 14:00'te → on dörtte, 4'ü → dördü, SGK'ya / SGK'ye, TL'lik → te lelik | liralık, 100'de (yüzde, a homograph of percent).
• Call centers are where Turkish TTS is deployed. A single wrong IBAN digit is a failed call. The public Turkish leaderboards score 495 short, digit-free everyday sentences, which cannot show this.
• Scores on the existing boards are dominated by measurement artefacts. Python str.lower() turns İ into i̇ and splits İstanbul into two words; spoken references that still contain digits reward systems that skip numbers; the Whisper language model "repairs" and rewrites. At ~1 % WER the top systems are a statistical tie. This benchmark fixes the normalizer, ships spoken-form multi-reference lattices, and reports confidence intervals and tie groups.

Composition

The pool was generated to the quota in generation/allocation.json (taxonomy §5), filtered by schema/QA checks, exact and near-duplicate removal and train/test decontamination, then split. Segment length is counted in spoken words (1.250.000 TL is 2 written tokens but 6 spoken words).

Setting × length

setting	short (1–5)	medium (6–20)	long (21–45)	very long (46+)	total	public	hidden
call_center	556	2080	958	412	4006	3069	937
general	388	1267	608	260	2523	1931	592
total	944	3347	1566	672	6529	5000	1529

Phenomenon groups

group	name	call-center	general	public	hidden	total
G01	Cardinal numbers	218	110	251	77	328
G02	Number morphology (suffixes on numerals, symbols and units)	218	142	275	85	360
G03	Ordinals, Roman numerals, enumerators	110	186	226	70	296
G04	Decimals, fractions, percent	208	121	251	78	329
G05	Currency & amounts	341	88	328	101	429
G06	Dates & years	219	110	252	77	329
G07	Times, durations, ranges	216	110	250	76	326
G08	Phone numbers & IVR keypad	231	28	199	60	259
G09	Identifiers	360	33	300	93	393
G10	Plates, addresses, postal codes	195	33	175	53	228
G11	Abbreviations, acronyms, letter spelling	260	163	324	99	423
G12	Units & measurements	110	152	202	60	262
G13	Symbols, math, scores, versions, web, LLM markup	120	208	251	77	328
G14	Foreign names, brands, code-switching	209	154	279	84	363
G15	Orthography & phonology	98	231	252	77	329
G16	Proper names	141	121	200	62	262
G17	Prosody & syntax	194	132	249	77	326
G18	Conversational	204	101	235	70	305
G19	Domain boilerplate	140	87	175	52	227
G20	Clean control (no TN triggers)	214	213	326	101	427

Domains

domain	name	public	hidden	total
CC01	Banking: cards, accounts, loans	486	160	646
CC02	Collections / tahsilat	278	82	360
CC03	Insurance: kasko, trafik, sağlık/TSS, DASK	277	84	361
CC04	Telecom / GSM / ISP	301	93	394
CC05	E-commerce & cargo	281	83	364
CC06	Energy & utilities: elektrik, doğalgaz, su	225	69	294
CC07	Healthcare appointments: MHRS, hospitals	217	71	288
CC08	Airline / travel / hotel	227	72	299
CC09	Automotive service / rent-a-car / ekspertiz	153	41	194
CC10	Public services: e-Devlet, SGK, tax, municipality	200	59	259
CC11	Retail & loyalty	147	47	194
CC12	Tech support	153	41	194
CC13	Real estate	81	16	97
CC14	Education	43	19	62
GN01	News: politics, society	329	112	441
GN02	Economy and markets news	65	19	84
GN03	Everyday conversation	274	81	355
GN04	Narrative and literature-like	215	62	277
GN05	Technology	184	60	244
GN06	Sports	184	58	242
GN07	Culture, history, arts	170	53	223
GN08	Health information	162	47	209
GN09	Weather and traffic announcements	155	42	197
GN10	Navigation	95	31	126
GN11	Science and education	98	27	125

Difficulty and WER visibility

difficulty	public	hidden	total	share
D1	682	202	884	13.5%
D2	2134	679	2813	43.1%
D3	2022	611	2633	40.3%
D4	162	37	199	3.0%

wer_visible	public	hidden	total	share
W	4119	1276	5395	82.6%
P	724	205	929	14.2%
N	157	48	205	3.1%

• D1 routine · D2 needs context or morphology · D3 hard (context disambiguation, rare conventions) · D4 diagnostic: genuinely ambiguous or convention-dependent, excluded from the headline.
• W errors show up as word errors · P partly WER-visible · N WER-blind (circumflex, stress minimal pairs, intonation, fillers): excluded from the headline, reported through human / pronunciation probes.

flag	public	hidden	total
headline	4549	1401	5950
robustness	69	16	85
multi-reference (n_paths > 1)	2525	803	3328

Subsets (inside public)

subset	n	call-center	general	short	medium	long	very long	headline
ci_smoke	500	308	192	72	256	118	54	458
adjudicated_core	1000	615	385	143	514	241	102	915
human_ceiling	600	370	230	95	390	115	0	577

• ci_smoke: fixed stratified subset for per-checkpoint regression runs.
• adjudicated_core: stratified subset on which native annotators adjudicate STT-vs-reference disagreements (human-adjudicated WER, reported separately).
• human_ceiling: stratified subset (no D4, no robustness items, ≤ 30 words preferred) to be read by native speakers; all STTs are run on those recordings to measure the STT error floor.

Lengths

split	n	min	p10	median	p90	max	mean words	median chars	max chars	total chars
public	5000	1	5	13	46	96	18.9	78	599	608890
hidden	1529	1	5	13	46	80	18.9	79	608	186354
all	6529	1	5	13	46	96	18.9	78	608	795244

Fields

field	type	description
id	str	trtb-NNNNN, assigned after a seeded shuffle (ids carry no stratum information)
version	str	dataset version
split	str	public (this release) or hidden
setting, domain, domain_name	str	call_center / general; CC01–CC14, GN01–GN11
utterance_type	str	e.g. verification, news_report, ivr_prompt
group, group_name	str	phenomenon group G01–G20
phenomenon_primary, phenomena_secondary	str, list[str]	phenomenon ids (generation/phenomena.json)
difficulty	str	D1–D4
wer_visible	str	W, P, N
robustness	bool	intentionally non-standard input (missing circumflex, wrong suffix, all caps, LLM markup)
headline	bool	difficulty != D4 and wer_visible != N and not robustness
text	str	written input, exactly what a production system would send (track R input)
text_spoken	str	canonical reading in Turkish words (track S input)
ref_lattice	str	normalized reference lattice, a b {c d|e} f; the first option of each group is canonical
ref_canonical	str	normalized canonical reading (tr_norm tokens)
n_paths	int	number of paths through the lattice
n_words, n_chars	int	spoken words of the canonical reading; characters of text
length_bucket	str	short 1–5, medium 6–20, long 21–45, very_long 46+ spoken words
alt_spans	list	{span, alts}: canonical substring of text_spoken and its accepted alternative readings (entity spans for entity accuracy)
acronyms	list	{written, type (L letter-wise / W word), canonical, alts} for every acronym in text
pron_hint	str	expected pronunciation of foreign words / brands (vatsap) for human or phoneme evaluation
notes	str	ambiguity notes for annotators
subsets	list[str]	ci_smoke, adjudicated_core, human_ceiling
source	str	provenance (llm-generated+llm-verified; human-review: pending)
canary	str	benchmark canary string
bid	str	generation batch item id (provenance)

Files: data/public/segments.parquet (viewer), segments.jsonl (canonical; its sha256 goes into every run manifest), segments.csv (lists JSON-encoded), contamination.jsonl (per-item 5-gram overlap with our training texts), hidden_commitments.jsonl, STATS.md (full build statistics), lexicon/ (acronym, unit and symbol lexicons used on both sides of scoring), code/ (normalizer and scorer snapshot), CANARY.txt.

Tracks

• Track R (raw, headline). The system receives text and must do its own text normalization. This is what a customer hears.
• Track S (spoken). The system receives text_spoken, which isolates acoustics and pronunciation from TN.
• TN penalty = WER_R − WER_S, reported per system and per phenomenon group.

Scoring protocol

1. Synthesis. One pre-registered default voice per system (or the fixed eight-speaker reference panel, assigned round-robin, for cloning systems); vendor default settings including their own text normalization; 1 take on all splits and 3 takes on adjudicated_core (seed 100000·s + i), pooled per segment; failures after 2 retries count as empty audio (all deletions) and the failure rate is reported. Every system row records the model id/commit or API model string and date, voice, settings and hardware.
2. Channels. fb (16 kHz), nb8k (8 kHz down → 16 kHz up; the protocol commonly used on Turkish TTS leaderboards, used for the comparability column) and g711 (300–3400 Hz band-pass, 8 kHz, A-law encode/decode).
3. STT panel. Primary whisper-large-v3-pdcompat: faster-whisper (pinned version and CTranslate2 revision of Systran/faster-whisper-large-v3, fp16) called exactly like the common Turkish TTS leaderboard harness — language="tr", beam_size=5, every other argument at its library default, non-batched (no VAD) — on the nb8k channel; fb is reported as a second column and a deterministic variant (temperature 0, no fallback, condition_on_previous_text=False) is kept as an ablation. Plus two open-weight STT models of different architectures (an LLM-based ASR and an acoustic CTC model). Results are reported per STT and as a panel mean; STT disagreement is reported as a reliability signal. Our in-house STT is diagnostic only and never enters the mean.
4. Normalization (tr_norm v1.0.0). Identical on both sides: NFC → protect Latin-I acronyms (IBAN, PIN) → Turkish casefold (İ→i, I→ı, stray U+0307 repaired) before lowercasing → glue apostrophe suffixes (Ankara'ya → ankaraya) → hyphen/slash → space → fold â/î/û → punctuation → space. The hypothesis is first verbalized by a Turkish TN (thousands ., decimal ,, times, ordinals, %, currency, units, the shipped acronym lexicon; when several readings exist, the one whose vowel harmony matches the written suffix wins). Whisper's English/Basic normalizers are never used.
5. Alignment. Track R: lattice-aware edit distance, minimum over all lattice paths (no path cap). Track S: strict against ref_canonical, lattice WER also reported.
6. Metrics. Corpus (micro) WER, CER (no spaces), space-insensitive CER, SER, insertion / deletion / substitution split, entity accuracy (exact match of alt_spans and numeric spans: amounts, dates, IBAN, phone, ids), failure rate, hallucination flag (hyp words > 1.5 × ref + 3) and truncation flag (hyp < 0.5 × ref). Headline numbers use only headline == true rows; D4, N and robustness rows are reported as diagnostics.
7. Statistics. Paired bootstrap over segments (B = 10,000, joint resampling across systems), 95 % CIs on every WER and every pairwise Δ, Holm-corrected pairwise tests, rank ranges and tie groups. There is no composite score: accuracy, naturalness (human), latency (TTFA p50/p95 at the client socket), throughput under concurrency and cost are separate columns.
8. Comparability column. For readers of existing Turkish TTS leaderboards, we also report Whisper large-v3 on nb8k with their score normalizer, and calibrate the harness by re-running their public test set for a few systems.

Hidden split, rotation and canary

• Hidden split. 1529 segments drawn from the same generator and strata are withheld. For each hidden row we publish sha256(salt + canonical JSON row) (hidden_commitments.jsonl, scheme: json.dumps(row, ensure_ascii=False, sort_keys=True, separators=(',', ':'))). Leaderboard rows report public and hidden scores side by side; a public − hidden gap larger than its CI flags possible over-fitting.
• Rotation. The hidden split is retired on a fixed schedule (target: quarterly). At retirement the salt and the rows are revealed, anyone can verify the commitments, the retired rows join a public archive and a fresh hidden split is generated with a new secret seed.
• Canary and decontamination. 6,554 candidate items were checked against 1,066,019 training sentences (sha256 60bdac347159dc1d…): 22 shared an 8-gram and 3 matched a training sentence exactly; all were dropped. Kept items: 970 share at least one 5-gram (mean 5-gram overlap 0.0158). Our own training-data builders reject any text that contains the canary or shares an 8-gram with public ∪ hidden (scripts/contamination.py guard, run in CI). We publish the check code and the sha256 of our training-text union, not the training text itself.

Limitations

• Text is LLM-generated and LLM-verified; human review is pending. Items were written by Claude Haiku and independently reviewed and corrected by Claude Opus (Claude Sonnet for the four groups without normalization triggers); the first three batches were written by Claude Sonnet. Model choice was made with blind audits (a quantized local open-weight LLM was tested and rejected: about 20 % unusable items). Every item passed an automatic validator (no digits or symbols in the spoken form, vowel-harmony checks on written suffixes, lattice ≤ 256 paths, length buckets). A blind audit of a stratified sample of 200 public items by two independent Claude Opus judges found 2.5 % [95 % CI 1.1–5.7] items unusable as references (group-weighted 2.3 %) and 19 % with any flaw, mostly casing/apostrophe consistency and tags; the items found were fixed. Native-speaker review of the canonical readings and alternatives has not been done yet; until then some references will be wrong or too strict. Report suspected errors (see Contact) and expect a corrected point release.
• Some phenomena are invisible to WER (circumflex/vowel length, stress minimal pairs, question intonation, palatal consonants, fillers). They are tagged wer_visible = N, excluded from the headline and need human or pronunciation-probe evaluation.
• STT bias. STT error at this level is the same order as TTS error. Whisper's language model can mask a mispronunciation or rewrite a correct one (mutfağı → mutfağa, ITN to digits). The STT panel, the human-ceiling recordings and the human-adjudicated core exist to bound this, but every automatic WER remains a proxy for intelligibility.
• Synthetic distribution. Sentences follow a designed quota, not real traffic frequencies, and use fictional names, invalid-checksum identifiers and random phone numbers.
• Naturalness is not measured by this dataset. Automatic MOS predictors are not validated for Turkish and are diagnostic only; human naturalness tests are run separately.

Conflict of interest

TR-TTS-Bench is built by Mank Yazılım, which develops and sells a Turkish TTS system (avaz3) that is evaluated on this benchmark. To limit the conflict: the data, normalizer and scorer are public; per-clip transcripts and run manifests will be published with every leaderboard run; the voice and settings of every system are pre-registered; our in-house STT never enters the STT mean; human evaluations will use an external rater pool; and we invite third-party re-runs, which we will list next to our own numbers.

Citation

@misc{trttsbench2026,
  title        = {TR-TTS-Bench: an open Turkish text-to-speech benchmark for text normalization and call-center speech},
  author       = {{Mank Yazılım}},
  year         = {2026},
  note         = {Version 0.1.0},
  howpublished = {Hugging Face dataset card}
}

Contact

Questions, reference corrections and third-party results: please open a discussion in the Community tab of this repository.

License

Text, references and annotations: CC BY 4.0. Synthesized audio of third-party systems is published only where the system's licence and terms of service allow it; otherwise only transcripts and metrics are published.