nineninesix/multilingual-tts-benchmark

Multilingual Speech Benchmark for Zero-Shot TTS

A voice-cloning and intelligibility benchmark for six language variants, built from Common Voice 17.0 by coverage-driven selection rather than random sampling. Every example pairs a reference clip of one speaker with a target text that speaker never read, so a system is asked to clone a voice and produce new speech, which is what zero-shot TTS is actually for.

Pipeline source code: https://github.com/nineninesix-ai/make-speech-benchmark — every number in this card is reproducible from it. Card generated at pipeline revision cfe006d.

Version 2.0. The audio, the row composition and the utt identifiers are unchanged from v1: anything already synthesised against v1 still joins. What changed is the measurement and the description of it — see What changed in v2.0.

---

Contents

• What this measures, and what it does not
• Quick start
• What changed in v2.0
• Subsets
• Human anchors — intelligibility
• Human anchors — speaker similarity
• Can a model beat the anchor?
• Evaluation protocol
• How to read the numbers
• Source data
• Quality control
• Example selection
• Data fields
• Reports
• Limitations
• Reproduction
• Personal and sensitive information
• Licence, citation, contact

---

What this measures, and what it does not

This is an intelligibility and voice-cloning benchmark, not a TTS quality benchmark. It measures two things:

• WER / CER on a recogniser reading the synthesis — a proxy for intelligibility;
• SIM — cosine similarity of speaker embeddings between the synthesis and the reference clip — a proxy for speaker identity.

It measures no naturalness axis at all: no subjective MOS or CMOS, and no predicted naturalness on the synthesis. A system can score 3 % WER and 0.95 SIM and still sound robotic, and nothing here would notice. It also contains no digits, no abbreviations and no long-form text — Common Voice sentences of 3-16 words — so text normalisation, the most common production TTS failure, is untested. Treat a good score as a necessary condition, not a sufficient one.

Quick start

from datasets import load_dataset

ds = load_dataset("nineninesix/multilingual-speech-benchmark", "en-US", split="main")
row = ds[0]
row["prompt_audio"]   # reference clip of the speaker, 16 kHz
row["text"]           # the target text, which this speaker never read
row["gt_audio"]       # a real recording of that text, by someone else
row["sim_ref_audio"]  # a second clip of the PROMPT speaker -> the SIM anchor

row["anchor_wer"]           # what the recogniser scores on the HUMAN recording
row["anchor_sim_wavlm_ft"]  # what two recordings of the same person score

Those last two are the reference points your model's numbers are read against, and they ship per row — along with the edit counts, so the corpus-level headline is derivable without leaving the dataset (see Data fields). The measurements behind them — S/D/I counts, transcripts, every recogniser and encoder — are browsable as separate configs:

wer = load_dataset("nineninesix/multilingual-speech-benchmark", "metrics-wer", split="train")
sim = load_dataset("nineninesix/multilingual-speech-benchmark", "metrics-sim", split="train")

Synthesise row["text"] conditioned on row["prompt_audio"], save it as {utt}.wav, then score:

git clone https://github.com/nineninesix-ai/make-speech-benchmark && cd make-speech-benchmark
uv venv && source .venv/bin/activate && uv pip install -e ".[all]"

msbench-wer --lang en-US --audio synth --asr whisper \
    --synth-dir out/mymodel/en-US --model-name mymodel
msbench-sim --lang en-US --mode synth --encoder wavlm_ft \
    --synth-dir out/mymodel/en-US --model-name mymodel
msbench-report --all

seed-tts-eval compatibility

The layout is consumable by the unmodified seed-tts-eval scripts, but read Human anchors — speaker similarity first: those scripts score SIM with wavlm_large_finetune.pth, whose values live near 0.65-0.69 here, while v1 published anchors from microsoft/wavlm-base-plus-sv, which live near 0.93. Comparing one against the other is the single easiest way to misread this benchmark, and v2 ships anchors for both.

What changed in v2.0

No data changed. Audio in all three streams is byte-identical, rows are in the same order, utt is stable. Verified by msbench.build.validate, which is run before every release.

Published numbers

The headline WER changes because the aggregation changed, not because the recogniser or the audio did. v1 averaged per-utterance error rates; v2 reports the corpus-level rate that seed-tts-eval uses, Σ(S+D+I) / Σ N_ref, and keeps the old form beside it labelled as legacy. On texts with a median of 6-7 words the two differ materially: one wrong word in a four-word sentence contributes 25 % under the macro mean regardless of its weight in the corpus.

Each column below isolates one change, measured on identical audio:

lang	published v1 (macro)	legacy (macro)	drift (env)	D-02 penalty	D-10 chunking	D-01 aggregation	v2 (corpus)	v2 corpus 95% CI
en-US	0.0806	0.0805	-0.0001	0.0002	0	-0.0033	0.0774	[0.0703, 0.0842]
es-ES	0.0485	0.0485	0	-0.0001	0	-0.0032	0.0452	[0.0397, 0.0512]
es-MX	0.0654	0.0652	-0.0002	-0.0004	0	-0.0035	0.0613	[0.0541, 0.0691]
nl-NL	0.0407	0.0407	0	0.0001	0	-0.0023	0.0385	[0.0331, 0.0453]
pt-BR	0.0823	0.0816	-0.0007	0.0001	0	-0.008	0.0737	[0.0608, 0.0884]
ky	0.1104	0.1104	0	0	0	-0.0106	0.0998	[0.0773, 0.1311]

• drift (env) — replaying v1's exact decoding parameters on current library and hardware versions. Worst case −0.0007; four subsets reproduce exactly. This is not a methodological change and is reported only so it is not mistaken for one.
• D-02 penalty — v1 decoded with repetition_penalty=1.1, which is not standard for WER evaluation and suppresses exactly the token pattern that autoregressive TTS failure produces. Removing it moves the human anchor by at most 0.0004, which is expected: human speech rarely loops. Whether it was masking model failures is a claim about synthesis and is not tested here.
• D-10 chunking — chunk_length_s=30 engaged Whisper's long-form path on clips of 2-6 s. Effect: exactly zero, transcripts 100 % identical. Removed as hygiene.
• D-01 aggregation — the entire remaining change.

Schema

change	detail
qc_hf_energy_db	new name for qc_bandwidth_hz, which holds decibels, not hertz: it is the share of energy above 5 kHz in dB. The old column is kept as a deprecated copy for one release
tags	was list<null>, a type that cannot hold a value; now list<string>
sim_ref_dur_bin	duration stratum of the second clip, so SIM breakdowns can control for it
gt_qc_*, simref_qc_*	9 columns each: DNSMOS and the QC battery on ground-truth and second-clip audio, which never faced the prompt-acceptance filters. Reporting only — nothing is filtered
phones (ky)	espeak dental markers t[ / d[ (1,584 and 1,132 occurrences) replaced by IPA t̪ / d̪. n_phones is unaffected
removed	qc_snr_db and qc_asr_cer were declared in v1 and 100 % null in every subset — never computed

New measurements

• a second recogniser on all six subsets (facebook/mms-1b-all, CTC without a language model) and a third encoder family for SIM;
• an impostor floor for SIM, which v1 lacks entirely;
• 95 % confidence intervals on every headline number, and n in every cell;
• per-utterance parquet artifacts shipped inside this repository, so any number here can be recomputed, sliced or re-tested without a GPU.

Corrections to the v1 card

claim in v1	corrected
second clip missing for "26-39 %" of rows	present for 60.6-87.9 %; see Subsets
SIM rises monotonically with prompt duration in five subsets	three under v1's encoder — en-US, es-MX, nl-NL
every example carries ground-truth audio and a second clip	GT missing for 7 examples; second clip present for 60.6-87.9 %
"Speakers"	speakers used, which is not speakers available in the source slice
anchors are a "physical ceiling"	human anchor. A model can exceed it — see Can a model beat the anchor?

Subsets

subset	examples	speakers used	with GT audio	with second clip
en-US	1500	1162	1500/1500	909/1500
es-ES	1500	722	1499/1500	1222/1500
es-MX	1500	929	1500/1500	1201/1500
nl-NL	1500	469	1500/1500	1319/1500
pt-BR	1500	247	1500/1500	983/1500
ky	700	181	694/700	599/700

Human anchors — intelligibility

Every figure below is a human anchor: a recogniser reading a real human recording of the target text. Without it a model's WER cannot be read — is 8 % on en-US a weak model, or roughly what the recogniser scores on human speech?

Headline aggregation is corpus-level, Σ(S+D+I) / Σ N_ref. Intervals are 95 % cluster bootstraps resampling speakers, not rows.

lang	ASR	n	speakers	WER corpus	95% CI	WER macro (v1)	CER corpus	exact	catastrophic
en-US	whisper	1500	1012	0.0774	[0.0703, 0.0842]	0.0807	0.0296	59.7%	1.4%
es-ES	whisper	1499	574	0.0452	[0.0397, 0.0512]	0.0484	0.0154	73.0%	0.9%
es-MX	whisper	1500	639	0.0613	[0.0541, 0.0691]	0.0648	0.0214	68.3%	1.3%
nl-NL	whisper	1500	236	0.0385	[0.0331, 0.0453]	0.0408	0.0108	76.0%	0.5%
pt-BR	whisper	1500	229	0.0737	[0.0608, 0.0884]	0.0817	0.0234	71.5%	3.5%
ky	gigaam	694	89	0.0998	[0.0773, 0.1311]	0.1104	0.0293	66.4%	5.5%

exact is the share of utterances transcribed with zero errors. catastrophic is the share above 50 % WER — the indicator that catches looping, babbling and dropped clauses long before they move the mean.

Three recognisers, and why the gaps matter

Whisper decodes with a strong internal language model, so a mispronounced or half-swallowed word is often repaired into the word the sentence implies; a system is then credited with intelligibility it did not produce. MMS is CTC with greedy decoding and no language model: it emits what it heard. Its absolute WER is higher everywhere, which is not a defect. Scribe is a second strong-LM read, from a different vendor and training set.

lang	A	B	WER A	WER B	delta (A−B)	95% CI	identical transcripts	n shared
en-US	elevenlabs	whisper	0.0515	0.0774	-0.0259	[-0.0309, -0.0209]	68.2%	1500
en-US	elevenlabs	mms	0.0515	0.1784	-0.1269	[-0.1357, -0.1183]	32.1%	1500
en-US	whisper	mms	0.0774	0.1784	-0.1011	[-0.1086, -0.0934]	32.0%	1500
es-ES	elevenlabs	whisper	0.0269	0.0452	-0.0183	[-0.0227, -0.0143]	78.8%	1499
es-ES	elevenlabs	mms	0.0269	0.1248	-0.0979	[-0.1065, -0.0895]	49.0%	1499
es-ES	whisper	mms	0.0452	0.1248	-0.0796	[-0.0871, -0.0717]	50.0%	1499
es-MX	elevenlabs	whisper	0.0351	0.0613	-0.0262	[-0.0321, -0.0207]	73.9%	1500
es-MX	elevenlabs	mms	0.0351	0.1482	-0.1131	[-0.1240, -0.1028]	42.0%	1500
es-MX	whisper	mms	0.0613	0.1482	-0.0869	[-0.0952, -0.0788]	43.0%	1500
nl-NL	elevenlabs	whisper	0.0159	0.0385	-0.0226	[-0.0277, -0.0186]	78.9%	1500
nl-NL	elevenlabs	mms	0.0159	0.078	-0.0621	[-0.0730, -0.0540]	58.7%	1500
nl-NL	whisper	mms	0.0385	0.078	-0.0395	[-0.0478, -0.0328]	57.3%	1500
pt-BR	elevenlabs	whisper	0.0531	0.0737	-0.0207	[-0.0307, -0.0120]	73.7%	1500
pt-BR	elevenlabs	mms	0.0531	0.2242	-0.1712	[-0.1953, -0.1497]	36.2%	1500
pt-BR	whisper	mms	0.0737	0.2242	-0.1505	[-0.1682, -0.1333]	35.4%	1500
ky	elevenlabs	mms	0.1829	0.2425	-0.0596	[-0.0838, -0.0361]	20.3%	694
ky	elevenlabs	gigaam	0.1829	0.0998	0.0831	[0.0629, 0.1046]	43.1%	694
ky	mms	gigaam	0.2425	0.0998	0.1427	[0.1273, 0.1593]	24.8%	694

Deltas are paired bootstraps over the utterances both recognisers scored; every interval above excludes zero.

Two of these gaps mean different things.

Whisper − MMS measures how much the intelligible reading depends on the listener's expectations. A large gap means the acoustics alone do not carry the sentence.

Scribe − Whisper measures something the benchmark could not see with one recogniser: how much of the "human anchor" was never the human at all. Scribe reads the same recordings 27-59 % more accurately, and with a lower catastrophic rate, so it is not buying accuracy with hallucination. On nl-NL the human anchor falls from 0.0385 to 0.0159 with zero catastrophic utterances. Whatever a v1-style anchor attributed to "human speech is hard" was substantially Whisper's own error, and the ceiling a synthesis system is measured against is correspondingly higher.

ky is the exception and runs the other way: Scribe (0.1829) is well behind GigaAM (0.0998), paired delta +0.0831 [0.0629, 0.1046]. See below.

Is the Kyrgyz subset usable?

Kyrgyz is the one subset where the primary recogniser (GigaAM-Multilingual) is not independently validated for this purpose, and its v1 anchor of 11.04 % sat above the 10 % threshold that calls for a cross-check. v2 performs that cross-check with MMS, which covers Kyrgyz through its kir adapter.

MMS scores 2.03x-3.04x the strong recogniser across the five Whisper subsets. Kyrgyz sits at 2.43x — mid-range. GigaAM therefore behaves, relative to a language-model-free CTC baseline, exactly as Whisper does on the other five subsets, and the cross-check does not find it anomalous.

A third recogniser now confirms it from the other direction. Scribe covers kir — which Whisper's 100 languages do not — and it lands at 0.1829, well behind GigaAM's 0.0998 (paired delta +0.0831 [0.0629, 0.1046]) while beating MMS. So on the five subsets where a general-purpose hosted model is the most accurate reader available, it is; on Kyrgyz, the language-specific model wins by a wide margin. That is what a genuinely competent ky recogniser looks like, and it is the opposite of what a broken one would produce.

Scribe's Kyrgyz output also carries a signature worth knowing about. The references use exactly the 36 letters of the Kyrgyz alphabet; 6.2 % of Scribe transcripts contain letters from neighbouring Turkic languages — ғ, қ, ұ, ә, і (Kazakh), ҡ (Bashkir). Folding those to their Kyrgyz counterparts recovers only 0.0063 WER, 3.4 % of its errors, so the spelling is not what makes it worse. It is a symptom: rows containing a foreign letter average 0.59 WER against 0.17 elsewhere. When Scribe drifts into a neighbouring orthography it is usually losing the whole utterance, not just the spelling.

The real limitation of ky is a sampling problem: 694 rows come from only 89 distinct ground-truth speakers, and the anchor's interval is correspondingly wide, [0.0773, 0.1311]. Use it for measurement, treat small differences between systems on it with suspicion, and quote the interval.

Human anchors — speaker similarity

An absolute SIM value is meaningless without a floor. v1 published an anchor and no baseline, so "0.87 against a ceiling of 0.93" read like 94 % of the way there. Whether that is good depends entirely on what two different speakers score, and v1 never measured it. v2 does, over 5,000 random cross-speaker prompt pairs per language per encoder:

lang	encoder	anchor	95% CI	floor mean	floor p95	usable range	n rows	speakers	distinct values
en-US	wavlm_sv	0.9317	[0.9283, 0.9347]	0.6119	0.8707	0.3198	909	631	629
en-US	wavlm_ft	0.652	[0.6424, 0.6601]	0.0725	0.2483	0.5795	909	631	630
en-US	ecapa	0.6141	[0.6042, 0.6227]	0.0822	0.2447	0.5319	909	631	631
es-ES	wavlm_sv	0.9435	[0.9392, 0.9469]	0.7286	0.9245	0.2149	1222	590	587
es-ES	wavlm_ft	0.6888	[0.6791, 0.6980]	0.1302	0.3183	0.5586	1222	590	590
es-ES	ecapa	0.6505	[0.6405, 0.6599]	0.132	0.3206	0.5185	1222	590	589
es-MX	wavlm_sv	0.9456	[0.9432, 0.9480]	0.7315	0.932	0.2141	1201	771	768
es-MX	wavlm_ft	0.6882	[0.6811, 0.6956]	0.1685	0.3576	0.5197	1201	771	769
es-MX	ecapa	0.635	[0.6268, 0.6438]	0.122	0.299	0.5131	1201	771	769
nl-NL	wavlm_sv	0.9285	[0.9235, 0.9331]	0.7446	0.9215	0.184	1319	412	412
nl-NL	wavlm_ft	0.6708	[0.6619, 0.6797]	0.174	0.352	0.4969	1319	412	412
nl-NL	ecapa	0.6232	[0.6134, 0.6336]	0.1866	0.3666	0.4366	1319	412	412
pt-BR	wavlm_sv	0.9339	[0.9265, 0.9403]	0.765	0.9248	0.1689	983	161	161
pt-BR	wavlm_ft	0.6346	[0.6166, 0.6502]	0.1864	0.3754	0.4482	983	161	161
pt-BR	ecapa	0.5763	[0.5578, 0.5930]	0.1257	0.2993	0.4506	983	161	161
ky	wavlm_sv	0.9214	[0.9121, 0.9302]	0.7173	0.9279	0.204	599	155	155
ky	wavlm_ft	0.6069	[0.5849, 0.6267]	0.184	0.3837	0.4229	599	155	155
ky	ecapa	0.5593	[0.5371, 0.5811]	0.1484	0.3428	0.411	599	155	155

• anchor — cos(second clip, prompt clip): two different recordings of the same person.
• floor mean / p95 — cross-speaker pairs. p95 is the practical false-accept level: a system scoring there is being confused with strangers one time in twenty.
• usable range — anchor minus floor. This is the entire span in which a cloning system can distinguish itself.

Read the wavlm_sv rows carefully. That is v1's encoder, and its usable range is 0.17-0.32 on a scale that looks like it runs to 1.0. Worse, its impostor p95 reaches 0.87-0.93 — on Kyrgyz the p95 impostor (0.9279) is above the human anchor (0.9214), meaning more than 5 % of random cross-speaker pairs outscore two recordings of the same person. A bare number near 0.93 from this encoder carries very little information. wavlm_ft and ecapa keep 0.41-0.58 of usable range and separate speakers far more cleanly.

Report the normalised score, which is readable where a bare cosine is not:

sim_norm = (sim_o − sim_floor) / (sim_anchor − sim_floor)

0 means indistinguishable from an impostor; 1 means as close as two recordings of the same person. Per-language coefficients are in reports/csv/sim.csv.

Which encoder to use

encoder	model	use it for
wavlm_sv	microsoft/wavlm-base-plus-sv	continuity with v1 numbers only
wavlm_ft	wavlm_large_finetune.pth	comparability with the seed-tts-eval literature
ecapa	speechbrain/spkrec-ecapa-voxceleb	a reading from outside the WavLM family

The third exists because of a circularity in the build: prompt QC rejected clips lying more than μ−2σ from their speaker's centroid as measured by WavLM-SV, and the anchor was then measured with WavLM-SV. The pool was pre-selected for homogeneity in the metric's own space. wavlm_ft shares that backbone and inherits the blind spot; ECAPA-TDNN shares neither architecture, pretraining nor training corpus, so where its anchor tracks WavLM's the anchor is a property of the speakers, and where it does not, the selection is showing through.

SIM against prompt duration

Under v1's encoder, the relationship is monotonic in three subsets, not five:

lang	<3.7	3.7-4.5	>=4.5	monotonic
en-US	0.9228	0.9284	0.9367	True
es-ES	0.9321	0.9476	0.9456	False
es-MX	0.9377	0.9455	0.9492	True
ky	0.9167	0.9321	0.9138	False
nl-NL	0.925	0.9309	0.9367	True
pt-BR	0.9262	0.9491	0.9399	False

The choice of encoder changes the answer: under wavlm_ft and ecapa, es-ES becomes monotonic too and only ky and pt-BR peak in the middle bin. Full tables with n per cell are in reports/csv/sim_breakdown.csv.

The confounder v1 did not control

The table above bins on the prompt clip only. The other clip in the pair — sim_ref_audio, down to about 2 s — was uncontrolled, and it moves the anchor:

lang	3.5-4.5	<3.5	>=4.5
en-US	0.948	0.9268	0.9432
es-ES	0.9528	0.9413	0.9572
es-MX	0.9519	0.9432	0.9574
ky	0.911	0.9223	0.9174
nl-NL	0.9431	0.9279	0.8895
pt-BR	0.9451	0.9327	0.9376

sim_ref_dur_bin ships as a column in v2 so any SIM breakdown can control for it.

Effective sample size

The anchor is constant within a speaker by construction: every example of a speaker shares one prompt clip and one second clip. So n rows is not the number of measurements — en-US's 909 rows carry 629 distinct values, and ky's 599 rows carry 155. Every interval in this card is a cluster bootstrap over speakers for this reason. A row-level bootstrap would report intervals roughly √(rows per speaker) too narrow.

Can a model beat the anchor?

Yes, and models do. The Seed-TTS paper's Table 1 (zero-shot in-context learning) reports English SIM 0.762 against a human 0.730, and Mandarin WER 1.115 % against a human 1.254 %. The anchor is not a physical ceiling and this card does not call it one.

The reason is structural. Synthesis is conditioned on the prompt and inherits its recording channel, microphone and room; the anchor compares two different recordings of the person, made at different times. The comparison is asymmetric in the model's favour. Three further biases point in various directions:

• circularity — the QC filter and the v1 SIM metric share an embedding space (above);
• transcript quality — the WER anchor mixes recogniser error, reader error and unverified crowd-sourced transcripts. The ky subset contains clear cases where both recognisers agree with each other and disagree with the reference, i.e. the reference is wrong;
• training contamination — Whisper has plausibly seen Common Voice in training, which pushes the anchor the other way.

None of these are quantified. Treat the anchor as a reference point, not a bound.

Evaluation protocol

Every parameter that determines a number. v1 stated none of them.

Preprocessing parity

The anchors are measured on audio that the build put into a canonical state: mono, 16 kHz, Silero-VAD trimmed with a 50 ms pad, peak-normalised to −1 dBFS. Your synthesis must go through the same pipeline or the comparison is not like-for-like — trailing silence and level differences both shift SIM. msbench.audio.prepare() does this and the drivers apply it by default.

Resampling uses soxr HQ. This matters: v1 resampled with np.interp, linear interpolation with no anti-aliasing filter. References are natively 16 kHz and never took that path, but synthesis at 22.05 / 24 / 44.1 kHz always did and arrived at the speaker encoder carrying aliasing artifacts — a 15 kHz tone that must vanish at 16 kHz survives at −3.1 dBr with 93 % of the residue folded onto 1 kHz, in the middle of the speech band. Every model's SIM was depressed by this; the anchor was not.

Stored dataset audio is not re-trimmed, because re-running VAD with a different Silero version than the build used would move the anchor.

ASR

model            openai/whisper-large-v3        (ai-sage/GigaAM-Multilingual rev "ctc" for ky)
dtype            float16
num_beams        1                              greedy
max_new_tokens   200                            the runaway guard
repetition_penalty  not set                     removed in v2
chunk_length_s   not set                        short-form path
batch            8
second opinion   facebook/mms-1b-all, per-language adapter, CTC greedy, no LM

ElevenLabs Scribe

model               scribe_v2
language_code       per subset, ISO 639-3 (eng / spa / nld / por / kir) — never auto-detect
tag_audio_events    false     default is TRUE; its tags would be scored as words
diarize             false
timestamps_granularity  none
seed                0
temperature         0
no_verbatim         false     it strips filler words, which are real content here
keyterms            unset     it biases decoding toward supplied words
workers             12        measured: ~8.8 clips/s; 24 draws 429s
timeout             300 s

Two properties of this backend have no counterpart in the local recognisers and must be understood before its numbers are used.

It is not deterministic, even pinned. Identical requests return different transcripts. seed alone changes nothing; seed with temperature=0 narrows the spread a great deal but does not close it. Measured by running ky three times end to end: only 69.5 % of transcripts are identical across all three, yet corpus WER lands at 0.1829 / 0.1805 / 0.1809 — a range of 0.0024 against a bootstrap interval 0.078 wide. The noise is real per utterance and negligible in the aggregate. Audit a single row and you may not reproduce it; quote a headline and you will. (scribe_v1 is less stable still, which is why v2 is the default here — the version number is not the reason.)

It hallucinates on short clips. On a 1.1 s pt-BR utterance where Whisper scores 0.00 it returned an unrelated sentence. audio_duration_secs comes back matching the clip, so this is a decoder failure, not a transport one. Read its catastrophic_rate alongside its WER, never the WER alone.

There is no batch endpoint: the API takes one file per request, and "batch" in ElevenLabs' terms means asynchronous delivery by webhook, which does not raise throughput. Concurrency is the only lever.

Text normalisation

Applied identically to reference and hypothesis:

NFC -> lowercase -> expand digits (num2words, per language; none for ky)
    -> strip punctuation -> collapse whitespace
PUNCT = !"#$%&()*+,-./:;<=>?@[\]^_`{|}~«»„""''…–—¿¡

Apostrophes are kept (word-internal: don't, 's-Gravenhage). Diacritics are kept — año/ano and sé/se are different words, and stripping them would mask real errors. This is where the normaliser deliberately differs from whisper.normalizers, which is too aggressive for non-English.

Aggregation

wer_corpus  = Σ(S+D+I) / Σ N_ref     headline, matches seed-tts-eval
wer_macro   = mean per-utterance     v1 legacy, reported alongside
cer         counts the space as a character (jiwer default)
exact_match = share of utterances with WER == 0
catastrophic_rate = share with WER > 0.5

Rows whose normalised reference is empty are counted (n_empty_ref), not silently dropped. In this dataset that count is 0 in every subset — the defect existed in v1's code but never affected a published number.

How to read the numbers

Confidence intervals. Every headline figure carries a 95 % cluster bootstrap over speakers (2,000 replicates, seed 20260725). Observations are not independent: one prompt serves up to 7 examples, GT-speaker concentration is uncapped, and the SIM anchor is constant within a speaker. Resampling rows would understate the interval.

Claiming "A beats B". Do not compare two independent intervals — that throws away the fact that both systems were measured on the same sentences. Use msbench.stats.paired_bootstrap, which resamples the shared speakers once per replicate and applies that resample to both systems, and report the delta with its interval and P(A better).

Per-cell n. Every breakdown table in reports/ carries n and the number of distinct speakers behind it, and flags cells with fewer than 30 speakers. Do not read a trend off a thin cell.

Cross-language comparison is not supported. Anchor-normalised comparison is valid within a language only. Differences between anchors are a property of the recogniser, not of the languages, and for ky it is a different recogniser entirely.

Source data

Common Voice 17.0 (fsicoli/common_voice_17_0), CC0. Text and audio are crowd-sourced; speaker metadata is self-declared and often absent.

Regional variants are not locales

Common Voice has no es-MX, pt-BR or nl-NL locale. These variants live in free-text, self-declared accents / variant columns, the field is multi-valued, and the labels themselves contain commas inside parentheses ("España: Norte peninsular (Asturias, Castilla y León, Cantabria)"). A naive split(",") shreds them. Labels are split on commas at parenthesis depth zero and matched exactly — substring matching gives false positives, since the Caribbean Spanish label contains "Costa del golfo de México" and would be picked up by a contains("México") test.

es-ES means the Castilian norm: Norte peninsular plus Centro-Sur. Andalusian is excluded — seseo/ceceo places it closer to es-MX, and 82 % of that slice is one speaker. /θ/ is present in es-ES and absent from es-MX, which is the point of having both.

Quality control

Prompt candidates were rejected on criteria that speech restoration cannot fix: not mono, no speech found, clipping above 1e-3, energy above 5 kHz below −45 dB (the signature of upsampling from 8 kHz), speech ratio below 0.75 after trimming, more than one speech segment after merging gaps under 0.25 s, and duration outside 2-12 s. Noise metrics were deliberately not used as filters, because references are expected to go through a restoration model downstream.

Two things about the qc_* columns that v1 did not state:

• they are computed on the untrimmed clip, while prompt_audio ships trimmed. This is why qc_lead_sil and qc_trail_sil are non-zero on audio that has had its edge silence removed;
• they cover prompts only.

The last point matters for the WER anchor. gt_audio and sim_ref_audio got the same preprocessing but never faced the rejection filters, so the anchor includes recordings that would not have been accepted as prompts. v2 measures how many, and ships the result as gt_qc_* and simref_qc_* columns:

lang	audio	n	DNSMOS OVRL	OVRL p05	clipping %	narrowband %	low speech %	would fail prompt QC %
en-US	prompt	1500	2.761	1.971	0	0.27	0.27	0.53
en-US	gt	1500	2.766	1.968	0	1.47	0.93	2.4
en-US	sim_ref	909	2.752	2.028	0	0.22	0	0.22
es-ES	prompt	1500	2.766	2.011	0	1.33	0.2	1.53
es-ES	gt	1499	2.8	2.041	0	2.33	0.4	2.74
es-ES	sim_ref	1222	2.759	1.991	0	1.64	0.16	1.8
es-MX	prompt	1500	2.68	1.887	0	0.53	0	0.53
es-MX	gt	1500	2.662	1.834	0	2.73	1.8	4.27
es-MX	sim_ref	1201	2.672	1.861	0	1.25	0	1.25
nl-NL	prompt	1500	2.808	2.152	0	0.47	0	0.47
nl-NL	gt	1500	2.871	2.218	0	1.13	0.13	1.27
nl-NL	sim_ref	1319	2.804	2.082	0	1.59	0	1.59
pt-BR	prompt	1500	2.726	2.004	0	0.4	0	0.4
pt-BR	gt	1500	2.783	2.005	0	2.47	1.07	3.53
pt-BR	sim_ref	983	2.697	2.049	0	0.61	0.61	1.22
ky	prompt	700	2.718	1.859	0	1.71	0	1.71
ky	gt	694	2.749	1.924	0	1.87	0.58	2.31
ky	sim_ref	599	2.671	1.812	0	0.67	0	0.67

would fail prompt QC % applies the prompt thresholds to each stream. Prompts pass by construction. Nothing is filtered — removing rows would break utt stability with v1 — so a user who wants a clean subset applies their own threshold and says so.

Example selection

Prompts and targets are decoupled. Each speaker contributes one fixed reference clip for all of their examples, which keeps SIM variance down; the target text comes from a different recording, usually a different speaker. The model is therefore always asked for speech that does not exist.

Selection is constrained rather than random: a target length distribution per language, a per-speaker cap, a minimum female-voice share where the source allows one, and full phonetic coverage. All six subsets cover 100 % of the candidate pool's phoneme and diphone inventory.

Data fields

Three audio streams per example, all mono 16 kHz, VAD-trimmed, peak-normalised to −1 dBFS:

field	what it is
prompt_audio	the reference clip — condition your model on this
prompt_audio_orig	the same clip at its original sample rate
gt_audio	a real recording of text, usually by a different speaker. Basis of the WER anchor
sim_ref_audio	a second clip of the prompt speaker, different text. Basis of the SIM anchor

The human anchors, per row — the benchmark's central numbers, carried in the data itself so they are visible without downloading anything else:

field	what it is
anchor_asr	which recogniser produced the primary anchor: whisper, or gigaam for ky
anchor_wer, anchor_cer	that recogniser's per-utterance error rate on gt_audio
anchor_subs, anchor_dels, anchor_ins, anchor_n_ref_words	the edit counts behind it, and the reference length
anchor_cer_err, anchor_n_ref_chars	the same for characters
anchor_hyp	what it actually transcribed, normalised — so a row's WER can be understood rather than only read
anchor_wer_mms, anchor_cer_mms + counts	MMS, CTC without a language model
anchor_wer_scribe, anchor_cer_scribe + counts	ElevenLabs Scribe v2, and anchor_hyp_scribe
anchor_sim_wavlm_sv, anchor_sim_wavlm_ft, anchor_sim_ecapa	cos(second clip, prompt clip) under each encoder

NaN where the underlying audio is absent (has_gt or has_sim_ref false).

anchor_wer is not the headline number. It is a per-utterance rate, and averaging it gives the macro aggregation that v2 keeps only for continuity with v1. The headline is corpus-level, and the counts are shipped so it is derivable from the dataset alone:

import datasets
d = datasets.load_dataset("nineninesix/multilingual-speech-benchmark", "en-US", split="main").to_pandas()

wer_corpus = (d.anchor_subs + d.anchor_dels + d.anchor_ins).sum() \
             / d.anchor_n_ref_words.sum()      # 0.0774 — the headline
wer_macro  = d.anchor_wer.mean()               # 0.0807 — v1 legacy

The same holds for any slice: filter the rows first, then sum. Averaging anchor_wer over a slice silently switches you back to the macro form.

One further caution: the SIM anchor is constant within a speaker by construction, so those columns are not independent observations. Cluster by speaker_id before putting an interval on anything.

Identity: utt (stable join key), lang, subset, source, cv_version.

Prompt: prompt_text, prompt_dur (after trimming), prompt_sr_orig, prompt_dur_bin, prompt_cv_path, prompt_split_origin.

Speaker: speaker_id, speaker_gender (+ _source), speaker_age, speaker_accent_label (raw label, secondary accents included), speaker_n_in_subset.

Target text: text, text_norm (produced by msbench.normalize, not a placeholder), n_words, n_chars, len_bin, phones (IPA), n_phones, punct_type, text_cv_path, text_split_origin.

Anchor availability: has_gt, gt_dur, gt_speaker_id, gt_same_speaker, gt_gender, has_sim_ref, sim_ref_dur, sim_ref_dur_bin, sim_ref_text.

Prompt QC (qc_*, on the untrimmed clip): qc_vad_speech_ratio, qc_lead_sil, qc_trail_sil, qc_clip_rate, qc_hf_energy_db (dB above 5 kHz), qc_bandwidth_hz (deprecated alias of the previous), qc_dnsmos_ovrl, qc_dnsmos_sig, qc_dnsmos_bak, qc_spk_centroid_dist.

Ground-truth and second-clip QC (new in v2, reporting only): gt_qc_* and simref_qc_*, each with dnsmos_ovrl, dnsmos_sig, dnsmos_bak, clip_rate, hf_energy_db, vad_speech_ratio, lead_sil, trail_sil, fails_prompt_qc.

Reserved for a future hard set, currently constant: category (general), subcategory, tags (empty list<string>), difficulty (0), has_digit, has_abbrev, has_foreign (all false), notes.

Reports

reports/ is part of the release, not an afterthought.

path	contents
reports/results.md	all anchors with intervals, breakdowns with per-cell n
reports/attribution.md	the v1 → v2 ladder, one cause per delta
reports/summary.md, reports/coverage_*.md	per-subset composition and coverage
reports/quality_{prompt,gt,sim_ref}.md	DNSMOS and QC per audio stream
reports/csv/*.csv	every table above, machine-readable
reports/per_utterance/*.parquet	per-utt metrics for every recogniser and encoder, with S/D/I counts, reference lengths, both normalised strings and speaker_id
reports/per_utterance/*.json	the exact configuration of every run

The per-utterance artifacts are what let you re-aggregate any number in this card, audit outliers, run your own paired tests, or slice by any column — without a GPU and without re-running a recogniser. v1 shipped aggregate markdown only.

Limitations

1. No naturalness axis. Neither subjective nor predicted. A system can score well here and sound robotic.
2. No hard set. Common Voice contains 0 % texts with digits, so text normalisation is untested.
3. No long-form. Texts are 3-16 words; long-context prosody and stability are untested.
4. ky rests on 89 ground-truth speakers for 694 rows, with a correspondingly wide interval.
5. The WER anchor mixes three error sources — recogniser error, reader error, and unverified crowd transcripts — with no way to separate them.
6. Whisper has plausibly seen Common Voice, which flatters the anchor in the opposite direction to the previous point. Neither bias is quantified.
7. The QC filter and v1's SIM encoder share an embedding space. Use ecapa to see past it.
8. v1's SIM encoder has almost no usable range, and on ky its impostor p95 exceeds the human anchor.
9. Observations are not independent; always cluster by speaker.
10. Ground-truth and second-clip audio were never rejection-filtered; 1.2-4.3 % would fail the prompt criteria.
11. utt stability was prioritised over data cleanliness. Nothing is filtered on the new QC columns.
12. The macro→corpus change makes v2 numbers incomparable with v1's unless you use the wer_macro column, which is retained for exactly that purpose.
13. The repetition-penalty removal is justified theoretically, not empirically: its effect on synthesis has not been measured here.
14. Cross-language comparison of anchors is not supported.
15. Packaged audio was resampled to 16 kHz with linear interpolation at build time, before v2 fixed the evaluation path. That is frozen into the data and is one more reason not to compare absolute SIM across datasets.
16. Prompt durations are 2.5-5 s, shorter than the 3-20 s Seed-TTS uses, and SIM rises with prompt duration — so even same-encoder comparison with that literature is only partial.
17. Speaker metadata is self-declared and missing for a large share of speakers (44 % of ky gender).
18. pt-BR splits are broken upstream: 97 % of test sentences also appear in train, and 9,464 clips are duplicated between train and dev.
19. n_syllables was dropped rather than approximated — every heuristic broke on hiatus or diphthongs. Use n_phones / gt_dur for speech rate.
20. Scribe is a paid, hosted, non-deterministic recogniser. It is the most accurate reader of these recordings on five subsets, but it cannot be the primary anchor of an open benchmark: reproducing it costs money, needs network access, and lands within ~0.002 rather than exactly. Whisper and GigaAM stay primary for that reason, not because they are better.
21. DNSMOS is a reporting metric here, never a filter. Reference quality was measured against SIM and explains under 1.6 % of its variance, so it is not a confounder — but that was measured on the human anchor, where noise affects both embeddings and partly cancels.

Reproduction

git clone https://github.com/nineninesix-ai/make-speech-benchmark && cd make-speech-benchmark
uv venv && source .venv/bin/activate && uv pip install -e ".[all]"

msbench-fetch                  # download this dataset
bash scripts/runbook.sh all    # every number in this card
msbench-validate               # v2 differs from v1 only where claimed
pytest                         # the aggregation gate

Selection stages S1-S3 are not re-run and are not needed for any of the above; they require the 66 GB Common Voice corpus and would change utt.

Two documents in the repository carry more detail than fits here: docs/PROTOCOL.md is the full measurement contract, and docs/DEFECTS.md is the catalogue of all 27 v1 defects with what each one cost.

Personal and sensitive information

The audio is human speech from Common Voice, contributed under CC0 by volunteers who consented to public release. Speaker identifiers are truncated Common Voice client_id hashes and are not linkable to a person by this dataset alone. Demographic fields are self-declared and frequently absent. The recordings are nonetheless biometric voice data: they can be used to build speaker models, and this dataset exists to measure exactly that capability. Do not use them to impersonate the contributors.

Sentences come from Common Voice's own text corpora and may contain the usual errors of crowd-sourced transcription; several were identified during this work where the reference text disagrees with what was actually said.

Licence, citation, contact

CC0-1.0, matching Common Voice 17.0.

@misc{multilingual_speech_benchmark_2026,
  title  = {Multilingual Speech Benchmark for Zero-Shot TTS},
  author = {nineninesix},
  year   = {2026},
  note   = {Version 2.0},
  url    = {https://huggingface.co/datasets/nineninesix/multilingual-speech-benchmark}
}

Pipeline: https://github.com/nineninesix-ai/make-speech-benchmark. Issues and pull requests welcome — adding a language, a recogniser backend or a speaker encoder each touch one file.

References

• Common Voice 17.0 — https://commonvoice.mozilla.org
• Seed-TTS (arXiv:2406.02430) and seed-tts-eval, the source of the corpus-level WER convention and the wavlm_large_finetune.pth SIM checkpoint
• Whisper (arXiv:2212.04356)
• MMS (arXiv:2305.13516)
• WavLM (arXiv:2110.13900)
• ECAPA-TDNN (arXiv:2005.07143)
• DNSMOS P.835 (arXiv:2110.01763)
• Silero VAD — https://github.com/snakers4/silero-vad