When a 2026 benchmark asked neural text-to-speech (TTS) systems to stress the right word in two versions of the same sentence, Pair-Correct scores were near zero for nearly all of them, and the best result was 11.5% even with stress marked explicitly in the input (Turetzky et al., CAST benchmark 2026). The same pattern runs through the rest of the prosody research: in a Japanese audiobook study, pitch error roughly doubled between vocoder resynthesis and full TTS, an emotion classifier that scored 99.64% on human recordings scored 15.31% on synthesized ones, and closed-source models trailed real recordings by 0.93 points in long-form expressive richness. Overall naturalness scores now sit close to human recordings (see our synthetic speech MOS statistics), but melody, timing and emphasis are where the remaining gap lives. We aggregated data from the Journal of the Brazilian Computer Society, Interspeech and the ISCA Speech Synthesis Workshops, the Blizzard Challenge, ACL 2026 Findings, the Journal of the Acoustical Society of America, LabPhon, and more than a dozen arXiv papers.
TL;DR
- 95 of 100 prosody-evaluation studies from 2020 to 2024 analyzed pitch (F0), and F0 RMSE was the most-used metric (Galdino et al., JBCS 2025).
- Log F0 RMSE correlates with human prosody ratings at only r = 0.30 (Yang et al., ProsodyEval 2025).
- Pitch error (F0 RMSE) was 14.03 for the vocoder-only HiFi-GAN row versus 27.93 for the best full TTS model (Xin et al., audiobook TTS 2022).
- A baseline Cantonese TTS made 32.76% F0 frame errors on declarative questions versus 13.53% on statements (Bai et al., Interspeech 2022).
- Listeners heard only 21.83% of baseline-synthesized declarative questions as questions (Bai et al., Interspeech 2022).
- Pair-Correct stress placement topped out at 11.5% across all TTS systems tested, while a text-only model reached 76.1% (Turetzky et al., CAST 2026).
- Listeners recognized anger in 37% of human recordings but 7% of synthesized clips (Abilbekov et al., KazEmoTTS 2024).
- Emotion recognition on CosyVoice 2 output fell from 83.93% (real speech) to 34.57% when prosody had to come from text (Ge et al., 2026).
- Isolated-sentence ratings explained only about 12% of the variance in paragraph ratings (Clark et al., SSW10 2019).
- Closed-source long-form TTS averaged 3.42 in expressive richness versus 4.35 for real speech (Pan et al., SwanBench-Speech, ACL 2026 Findings).
- Human readers ran at 3.20 syllables per second when reading versus 4.96 when speaking freely (Jacewicz et al., JASA 2010).
1. Measuring Prosody: The Metrics Problem
The field measures prosody mostly with one number, and that number barely tracks what listeners hear. Log F0 RMSE, the default pitch-error metric, correlated with human prosody ratings at r = 0.30, less than half the correlation of a newer token-based metric (Yang et al., ProsodyEval 2025). F0 (fundamental frequency) is the acoustic measure behind perceived pitch, and RMSE (root mean square error) compares the synthetic pitch curve frame by frame with one human recording of the same sentence.
The problem is structural: a sentence can be read well with many different melodies, so a distance to one reference punishes valid variation. Even large audio language models are not yet a fix, since Gemini 2.5 Pro scored a correlation of r = 0.27 with human prosody ratings in the same benchmark of 1,000 samples, 7 TTS systems and 2,000 human ratings.
| Metric | Value | Source |
|---|---|---|
| Prosody-evaluation studies reviewed (published 2020 to 2024) | 100 | Galdino et al., JBCS 2025 |
| Studies that analyzed pitch (F0) | 95 of 100 | Galdino et al., JBCS 2025 |
| Studies by language: English / Mandarin Chinese / Japanese | 48 / 27 / 15 | Galdino et al., JBCS 2025 |
| Correlation with human prosody ratings: log F0 RMSE | r = 0.30 [0.19, 0.40] | Yang et al., ProsodyEval 2025 |
| Correlation with human prosody ratings: DS-WED (token-based metric) | r = 0.77 [0.73, 0.81] | Yang et al., ProsodyEval 2025 |
| Gemini 2.5 Pro as prosody judge, correlation with human ratings | r = 0.27 | Yang et al., ProsodyEval 2025 |
Context: the systematic review in the Journal of the Brazilian Computer Society also found that LJ Speech, a single-speaker English corpus of read non-fiction, was the most-used dataset, so a good part of the published evidence describes one reading style in one language. A separate benchmark of 35 TTS systems from 2008 to 2024 reports prosody mattering more for later systems, and its authors suggest human evaluators focus on it more as noise and intelligibility problems fade (Minixhofer et al., TTSDS 2024).
2. Measured Pitch and Duration Error Against Human Recordings
When a study separates the vocoder from the stage that predicts prosody from text, most of the pitch error appears in the second one. In a Japanese audiobook study, F0 RMSE was 14.03 for the vocoder-only HiFi-GAN row and 27.93 for the best full TTS model, about 2 times higher (27.93 / 14.03 = 1.99) (Xin et al., audiobook TTS 2022). A vocoder is the final stage that turns acoustic features into a waveform, so a vocoder-only score approximates the error floor of the audio chain itself.
Newer predictors shave only a few points off. A 2025 prosody model trained on about 750 hours of English audio matched the reference pitch in 43.9% of frames, up from 40.2% for the strongest baseline. Prosody variety is also fragile: preference fine-tuning (DPO) cut one system’s measured prosody diversity by 18.8%.
| Metric | Value | Source |
|---|---|---|
| F0 RMSE: vocoder only (HiFi-GAN) vs best full TTS model (Japanese audiobooks) | 14.03 vs 27.93 | Xin et al., audiobook TTS 2022 |
| Gross pitch error: vocoder only (HiFi-GAN) vs best TTS model vs baseline | 2.2 vs 12.57 vs 14.16 | Xin et al., audiobook TTS 2022 |
| AB preference: context-aware model vs baseline (40 listeners) | 61.5% vs 38.5% | Xin et al., audiobook TTS 2022 |
| F0 RMSE: hierarchical long-form model vs FastSpeech 2 with BERT baseline (Mandarin audiobooks) | 36.61 vs 51.12 | Guo et al., HiGNN-TTS 2023 |
| Duration MSE: hierarchical long-form model vs baseline | 0.1083 vs 0.1421 | Guo et al., HiGNN-TTS 2023 |
| Raw pitch accuracy (share of frames close to the human pitch): ProMode vs StyleTTS 2 baseline | 43.9% vs 40.2% | Eren et al., ProMode 2025 |
| Change in prosody diversity after DPO fine-tuning, CosyVoice 2 (LibriSpeech) | -18.8% | Yang et al., ProsodyEval 2025 |
Outlier note: F0 RMSE values are only comparable inside one paper, because units, alignment and datasets differ. The ProMode paper does not state units for its RMSE, and the HiGNN-TTS text describes its own F0 result in words that contradict its table, so we report the table values.
3. Questions, Syntax and Homographs: Reading the Sentence Type Right
Question intonation fails where the grammar gives no cue. A baseline Cantonese Tacotron 2 system made F0 frame errors on 32.76% of frames in declarative questions, against 13.53% in statements, and listeners identified only 21.83% of its declarative questions as questions (Bai et al., Interspeech 2022). A declarative question is a statement-shaped sentence that should end with a rising pitch; in the CanTTS corpus they are about 1,000 of 12,010 utterances, so the model rarely sees the pattern. F0 frame error counts frames where pitch, or the presence of voicing, departs clearly from the human recording.
The fix in that study, a separate question-versus-statement text classifier, lifted identification to 98.41%. Newer systems remain uneven: in the 2025 EmergentTTS-Eval benchmark, scored by an AI judge against a commercial baseline where 50% means parity, the open-source models Sesame1B, Orpheus-TTS and Qwen 2.5 Omni won 31.78%, 39.64% and 49.28% of question prompts (Manku et al., Boson AI, EmergentTTS-Eval 2025).
| Metric | Value | Source |
|---|---|---|
| Declarative questions identified as questions: baseline Tacotron 2 vs model with question classifier | 21.83% vs 98.41% | Bai et al., Interspeech 2022 |
| F0 frame error on declarative questions vs on statements, baseline Tacotron 2 | 32.76% vs 13.53% | Bai et al., Interspeech 2022 |
| Question intonation perceived correctly: vocoder resynthesis vs two emotional TTS baselines | 99.2% vs 81.6% and 85.9% | Tang et al., QI-TTS 2023 |
| Questions category win rate vs baseline: GPT-4o-audio (best overall) vs Sesame1B | 56.96% vs 31.78% | Boson AI, EmergentTTS-Eval 2025 |
| Syntactic complexity win rate vs baseline: Orpheus-TTS vs Sesame1B | 38.92% vs 18.88% | Boson AI, EmergentTTS-Eval 2025 |
| Complex pronunciation win rate vs baseline: Qwen 2.5 Omni | 3.30% | Boson AI, EmergentTTS-Eval 2025 |
| Agreement between AI judge ranking and human ranking (Spearman) | 90.5% | Boson AI, EmergentTTS-Eval 2025 |
| French homograph test (Blizzard 2023): chance level vs best systems | 50% vs more than 80% | Perrotin et al., Blizzard Challenge 2023 |
Context: homographs (words spelled the same but pronounced differently, such as French ‘est’ as a verb or a compass point) were scored by 218 listeners in an ABX test in the Blizzard Challenge 2023, and the organizers linked the best results to systems that used a large language model in the text encoder. The QI-TTS study dates from March 2023; most recent available data for that measure.
4. Emphasis and Contrastive Stress
Putting stress on the right word is the clearest current failure. In the CAST benchmark of 113 contrastive pairs (226 items), Pair-Correct scores (the right word stressed in both versions of a pair) were near zero for nearly all systems, and the best was 11.5%, reached only when stress was marked explicitly in the input (Turetzky et al., CAST 2026). The same benchmark showed that a text-only language model can work out the intended stress from context 76.1% of the time, so the information is available: TTS systems do not turn it into sound.
Giving the TTS model the preceding context brought no clear improvement for any system. Older listener studies point the same way: error marks on synthetic speech cluster around prosodic boundaries, and extra prosodic variation is often heard as a mistake when it does not fit the context.
| Metric | Value | Source |
|---|---|---|
| Target word stressed (Hit) / both sides correct (Pair-Correct): Kokoro | 38.1% / 0.0% | Turetzky et al., CAST 2026 |
| Hit / Pair-Correct: GPT-4o-mini-tts with explicit stress marking | 51.8% / 11.5% | Turetzky et al., CAST 2026 |
| Hit / Pair-Correct: text-only Claude Haiku 4.5 predicting stress from context | 88.1% / 76.1% | Turetzky et al., CAST 2026 |
| Emphasized word identified by listeners, EME-TTS: mean vs surprise emotion | 78% vs 64% | Li et al., EME-TTS 2025 |
| Correlation between prosody rating and listener-marked error rate | R = -0.75 | Gutierrez et al., SSW11 2021 |
| Share of times FastPitch’s most error-marked word came right before punctuation (audiobook text) | 73% | Gutierrez et al., SSW11 2021 |
Context: listener agreement on where stress falls in synthetic speech is itself modest (Fleiss’ kappa = 0.36 among three listeners in the same benchmark), which is one reason emphasis is hard to measure. In EME-TTS (11 listeners), surprise scored lowest, and the authors attribute this to the frequent pitch rise at the end of surprised speech, which led listeners to hear the final word as stressed. The rapid prosody transcription study is from 2021; most recent available data for listener error-marking.
5. Emotional Prosody: What Listeners and Classifiers Recognize
Emotion is carried largely by prosody, and synthetic emotion is recognized far less often than recorded emotion. In a Kazakh study with 64 listeners, the intended emotion was recognized in 43% of human recordings but 32% of synthesized clips, and anger dropped from 37% to 7% (Abilbekov et al., KazEmoTTS 2024).
A 2026 analysis traced where the emotion gets lost. On CosyVoice 2, emotion recognition stayed at 82.43% after vocoder reconstruction and 84.07% when the model was fed the real speech tokens, close to 83.93% for real speech, but fell to 34.57% when the model had to generate the speech from text (Ge et al., On the Emotion Understanding of Synthesized Speech, 2026). The loss sits in the text-to-prosody step, not in audio quality. For how recognition models perform on human voices, see our speech emotion recognition statistics.
| Metric | Value | Source |
|---|---|---|
| Listener emotion recognition: human recordings vs synthesized (6 emotions) | 43% vs 32% | Abilbekov et al., KazEmoTTS 2024 |
| Listener recognition of anger: human recordings vs synthesized | 37% vs 7% | Abilbekov et al., KazEmoTTS 2024 |
| Emotion classifier accuracy (emotion2vec): human TESS recordings vs synthesized TESS | 99.64% vs 15.31% | Ge et al., 2026 |
| Emotion classifier accuracy: human IEMOCAP vs synthesized IEMOCAP | 71.28% vs 49.58% | Ge et al., 2026 |
| CosyVoice 2 on TESS: real speech vs vocoder only vs text-driven generation | 83.93% vs 82.43% vs 34.57% | Ge et al., 2026 |
| Emotion classifier accuracy on gpt-4o-tts output (CREMA-D / TESS sentences) | 32.94% / 46.42% | Ge et al., 2026 |
| Emotion reclassification accuracy: vocoded human recordings vs best and weakest of three TTS systems | 0.931 vs 0.749 and 0.623 | Tang et al., ED-TTS 2024 |
| Emotions category win rate vs baseline: GPT-4o-audio vs Sesame1B | 88.84% vs 7.32% | Boson AI, EmergentTTS-Eval 2025 |
Outlier note: the 2026 classifier figures are machine scores, not listener judgments, and a classifier trained on synthesized speech reached 91.84% on synthesized TESS but 15.00% on human TESS, which suggests synthetic and human emotion differ acoustically. ED-TTS used its own emotion classifier, so its scale is not directly comparable to emotion2vec.
6. Long-Form Reading: Where Sentence Tests Stop Working
Prosody that sounds fine one sentence at a time can drift, flatten or misplace pauses across a paragraph. In a news-reading study, even a paragraph’s best single-sentence rating explained only about 12% of the variance in the rating of the whole paragraph (maximum sentence MOS vs paragraph MOS, r = 0.345, p < 0.01) (Clark et al., Evaluating Long-form Text-to-Speech, SSW10 2019). The authors concluded that sentence-level ratings are “bad predictors of paragraph MOS”.
The 2026 picture is narrower but not closed. On SwanBench-Speech, the best closed-source system came within 0.09 of real recordings on the prosody score (3.95 vs 4.04), but the closed-source average trailed real speech by 0.93 points on both expressive richness (emotional resonance, character portrayal, storytelling) and expressive hierarchy (paragraph-level expressive dynamics), and most models degraded once a passage passed 100 words (Pan et al., SwanBench-Speech, ACL 2026 Findings). For producers working with narrated training and explainer content, our AI voiceover statistics cover adoption, and an AI voice generator is easiest to judge on a full paragraph, not one line.
| Metric | Value | Source |
|---|---|---|
| Correlation, mean sentence rating vs paragraph rating (news) | r = 0.296 (p < 0.05) | Clark et al., SSW10 2019 |
| Mean rating, short sentences vs 40-second passages, best model | 4.09 vs 3.86 | Guo et al., HiGNN-TTS 2023 |
| Mean rating, short sentences vs 40-second passages, baseline | 3.52 vs 3.34 | Guo et al., HiGNN-TTS 2023 |
| Audiobook listening test: listeners / ratings / segment length | 204 / 33,750 / about 2 minutes | Zhang et al., SSW 2023 |
| Overall quality gain from long-form enhancements (abstract) | +0.47 MOS | Zhang et al., SSW 2023 |
| Comparative score after removing multi-sentence context: female / male voice | -0.42 / -0.58 | Zhang et al., SSW 2023 |
| Expressive richness: real speech vs closed-source average vs open-source average | 4.35 vs 3.42 vs 3.03 | Pan et al., SwanBench-Speech 2026 |
| Expressive hierarchy (paragraph-level dynamics): real speech vs closed-source average | 3.94 vs 3.01 | Pan et al., SwanBench-Speech 2026 |
Context: SwanBench scores are automatic. Its prosody metric agreed with 10 human evaluators at SRCC = 0.82 on 50 pairs, and its richness and hierarchy judge (Gemini 3 Pro) at 0.71 and 0.62. The audiobook synthesis paper states that its enhancements “reduce the huge gap between baseline synthesis and natural recordings” without closing it. The Clark et al. study is from 2019; most recent available data for the sentence-versus-paragraph comparison.
7. The Human Baseline: How Much Real Prosody Varies
Any “error” figure assumes a target, and human prosody is a wide target. Across 192 American English speakers, articulation rate was 3.20 syllables per second when reading sentences and 4.96 in spontaneous talk (Jacewicz et al., JASA 2010). Articulation rate counts syllables per second with pauses removed, so it isolates tempo from hesitation.
Variation inside one person is the bigger story: in the same data, phrase-to-phrase variation within a speaker exceeded variation between speakers. A corpus study presented at LabPhon 19, with nearly 120,000 utterances, found that individual speakers vary more than dialects do. This is why a single reference recording makes a weak yardstick, as section 1 showed for F0 RMSE.
| Metric | Value | Source |
|---|---|---|
| Articulation rate, read sentences vs spontaneous talk (syllables per second) | 3.20 vs 4.96 | Jacewicz et al., JASA 2010 |
| Spontaneous rate, Wisconsin vs North Carolina speakers (syllables per second) | 5.21 vs 4.80 | Jacewicz et al., JASA 2010 |
| Reading rate, young adults vs oldest adults (syllables per second) | 3.58 vs 3.0 | Jacewicz et al., JASA 2010 |
| Speaking-rate variance: within speaker vs between speakers | 1.0 vs 0.19 | Jacewicz et al., JASA 2010 |
| Corpus size: utterances / speakers / dialects | 119,736 / 1,115 / 15 | Tanner et al., LabPhon 19 |
| Estimated difference in articulation rate across English dialects | about 40% (about 2 syllables per second) | Tanner et al., LabPhon 19 |
| Probability that speaker variation exceeds dialect variation | 0.98 | Tanner et al., LabPhon 19 |
Context: the SPADE articulation-rate paper used 14 corpora of spontaneous speech from British and North American English. The Jacewicz data date from 2010; most recent available data for read versus spontaneous tempo in a single controlled sample. For another prosodic feature where listener judgments diverge from acoustics, see our vocal fry statistics.
Summary: Prosody and Intonation Error in Neural TTS by the Numbers
| Metric | Value | Source |
|---|---|---|
| Prosody studies (2020 to 2024) that analyzed pitch | 95 of 100 | Galdino et al., JBCS 2025 |
| Log F0 RMSE correlation with human prosody ratings | r = 0.30 | Yang et al., ProsodyEval 2025 |
| F0 RMSE: vocoder only vs best full TTS | 14.03 vs 27.93 | Xin et al., 2022 |
| Raw pitch accuracy of a 2025 prosody predictor | 43.9% | Eren et al., ProMode 2025 |
| F0 frame error, declarative questions vs statements (baseline) | 32.76% vs 13.53% | Bai et al., Interspeech 2022 |
| Declarative questions heard as questions: baseline vs classifier model | 21.83% vs 98.41% | Bai et al., Interspeech 2022 |
| Questions win rate vs baseline: GPT-4o-audio vs Sesame1B | 56.96% vs 31.78% | Boson AI, EmergentTTS-Eval 2025 |
| Homograph test: chance vs best systems | 50% vs more than 80% | Perrotin et al., Blizzard 2023 |
| Pair-Correct contrastive stress: best TTS vs text-only language model | 11.5% vs 76.1% | Turetzky et al., CAST 2026 |
| Listener emotion recognition: human vs synthesized | 43% vs 32% | Abilbekov et al., KazEmoTTS 2024 |
| Emotion classifier: human vs synthesized TESS | 99.64% vs 15.31% | Ge et al., 2026 |
| CosyVoice 2 emotion accuracy: real speech vs text-driven | 83.93% vs 34.57% | Ge et al., 2026 |
| Sentence rating vs paragraph rating, variance explained | about 12% | Clark et al., SSW10 2019 |
| Expressive richness: real speech vs closed-source average | 4.35 vs 3.42 | Pan et al., SwanBench-Speech 2026 |
| Short sentences vs 40-second passages, best model | 4.09 vs 3.86 | Guo et al., HiGNN-TTS 2023 |
| Read vs spontaneous articulation rate (syllables per second) | 3.20 vs 4.96 | Jacewicz et al., JASA 2010 |
| Articulation-rate difference across English dialects | about 40% | Tanner et al., LabPhon 19 |
Methodology and Sources
Every figure above was read in the paper, proceedings page or full text that reported it. Where a paper showed results only in a chart, we used its stated conclusions and dropped approximate values. Machine-judged scores (AI judges, emotion classifiers, automatic benchmarks) are labeled as such in the text and should not be read as listener percentages.
- Galdino, Matos, Svartman and Aluisio, “The evaluation of prosody in speech synthesis: a systematic review”, Journal of the Brazilian Computer Society, July 9, 2025
- Yang et al., “Measuring Prosody Diversity in Zero-Shot TTS” (ProsodyEval), arXiv 2509.19928, revised April 1, 2026
- Minixhofer, Klejch and Bell, “TTSDS - Text-to-Speech Distribution Score”, arXiv 2407.12707, 2024
- Xin et al., “Improving Speech Prosody of Audiobook Text-to-Speech Synthesis with Acoustic and Textual Contexts”, arXiv 2211.02336, 2022
- Guo et al., “HiGNN-TTS: Hierarchical Prosody Modeling with Graph Neural Networks for Expressive Long-form TTS”, arXiv 2309.13907, 2023
- Eren et al., “ProMode: A Speech Prosody Model Conditioned on Acoustic and Textual Inputs”, arXiv 2508.09389, 2025
- Bai, Ko and Zhang, “A Study of Modeling Rising Intonation in Cantonese Neural Speech Synthesis”, Interspeech 2022
- Tang et al., “QI-TTS: Questioning Intonation Control for Emotional Speech Synthesis”, arXiv 2303.07682, 2023
- Manku et al. (Boson AI), “EmergentTTS-Eval”, arXiv 2505.23009, May 29, 2025
- Perrotin, Stephenson, Gerber and Bailly, “The Blizzard Challenge 2023”, ISCA Archive
- Turetzky et al., “Knowing What to Stress: A Discourse-Conditioned Text-to-Speech Benchmark” (CAST), arXiv 2604.10580, April 12, 2026
- Li et al., EME-TTS, on the link between emphasis and emotion in speech synthesis, arXiv 2507.12015, 2025
- Gutierrez, Oplustil-Gallegos and Lai, “Location, Location”, ISCA SSW11, 2021
- Abilbekov et al., “KazEmoTTS: A Dataset for Kazakh Emotional Text-to-Speech Synthesis”, arXiv 2404.01033, 2024
- Ge et al., “On the Emotion Understanding of Synthesized Speech”, arXiv 2603.16483, March 17, 2026
- Tang et al., “ED-TTS: Multi-Scale Emotion Modeling using Cross-Domain Emotion Diarization”, arXiv 2401.08166, 2024
- Clark, Silen, Kenter and Leith, “Evaluating Long-form Text-to-Speech: Comparing the Ratings of Sentences and Paragraphs”, SSW10, arXiv 1909.03965, 2019
- Zhang et al., “Audiobook synthesis with long-form neural text-to-speech”, ISCA SSW12, 2023
- Pan et al., “Comprehensive Benchmarking of Long-Form Speech Generation in Diverse Scenarios” (SwanBench-Speech), ACL 2026 Findings, arXiv 2605.28618
- Jacewicz, Fox and Wei, “Between-speaker and within-speaker variation in speech tempo of American English”, Journal of the Acoustical Society of America, 2010
- Tanner et al. and the SPADE Data Consortium, “Exposing the anatomy of articulation rate across English dialects and speakers”, LabPhon 19
- Data watch: F0 RMSE values are not comparable across papers (different units, alignment and datasets), and the ProMode paper does not state units. Xin et al. list a HiFi-GAN row without describing it; we read it as vocoder-only output because its spectral distortion is far below the TTS rows, which is our inference. Jacewicz et al. give 5.21 and 4.80 syllables per second in the text but 5.38 and 4.97 in a table for the same dialects; we report the text. The SSW 2023 audiobook paper gives a 0.47 MOS gain in its abstract and 0.46 in its body text; its per-system chart values could not be mapped reliably, so we omitted them. SwanBench-Speech says the hierarchy gap is “over half a point” while its table averages imply 0.93; we report the table. EmergentTTS-Eval and SwanBench results come from AI judges validated against small human panels, and later technical reports that rewrote the EmergentTTS texts are not comparable with the original paper. The QI-TTS (2023), Xin et al. (2022), Bai et al. (2022), Gutierrez et al. (2021), Clark et al. (2019) and Jacewicz et al. (2010) figures are older than three years and are the most recent available data for those specific measures. Several 2026 sources are preprints and may change on peer review. The next Blizzard Challenge results and Speech Prosody proceedings may add newer question-intonation and long-form data.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the next Blizzard Challenge results are published on the ISCA Archive.