Listeners identified 95.4% of naturally spoken American English vowels in recordings from 139 talkers, against 94.4% in the classic 76-talker Peterson and Barney set (Hillenbrand et al., JASA 2011). Strip the same vowels down to formant-only synthesis and accuracy falls to 88.5%; reduce them to pure sine tones that only trace the formants and untrained listeners manage 55.0%. Perception of formants also matures slowly: Dutch children aged 6 to 11 needed an average 49% change in F1 before they noticed it, against 33% for adults (Polsterer et al., JASA Express Letters 2026). We aggregated data from the Journal of the Acoustical Society of America (Peterson and Barney, Hillenbrand, Vorperian, Chung, Chen), the Journal of Phonetics, the Journal of Communication Disorders, the VoxCommunis corpus (LREC 2022) and a 2026 JASA Express Letters study. If you came here from our singing formant and belting acoustics roundup, this page covers the speech side: the everyday vowel targets of ordinary talkers.
A formant is a resonance of the vocal tract, a band of frequencies the throat and mouth amplify. F1 rises as the jaw and tongue lower (open vowels), F2 rises as the tongue moves forward (front vowels), and F3 matters most for r-colored vowels such as the one in ‘heard’. Every table below uses hertz (Hz) unless stated otherwise.
TL;DR
- Peterson and Barney recorded 76 talkers (33 men, 28 women, 15 children) saying 10 vowels twice, 1,520 vowels in total (Praat manual, Peterson and Barney 1952 table).
- Natural vowels were identified correctly 94.4% (Peterson and Barney) and 95.4% (Hillenbrand) of the time (Hillenbrand et al., JASA 2011).
- Mean corner-vowel formants differ across published American English datasets by about 700 Hz for women’s F2 of /u/ (Kent and Vorperian, Journal of Communication Disorders 2018).
- American English /i/ F2: 2146 Hz in men, 2570 Hz in women, 2964 Hz in 5-year-olds, 3220 Hz in 2-year-olds (Chung et al., JASA 2012).
- An F2 sex difference already exists at age 4 for /i/ and /u/ (Vorperian et al., JASA 2019).
- A formant classifier trained on one US dialect fell from 88.9% to 44.8% correct on another (Jacewicz and Fox, JASA 2012).
- Formant movement inside a vowel was 477 Hz in speakers aged 51-65 vs 380 Hz in speakers aged 21-33 (Jacewicz, Fox and Salmons, Journal of Phonetics 2011).
- Average F1 of [i] and [u] correlates at r = 0.83 across 31 languages (Ahn and Chodroff, VoxCommunis, LREC 2022).
- Mean LPC F1 error reaches 58 Hz (9%) when F0 is above 300 Hz (Chen, Whalen and Shadle, JASA 2019).
- Even trained phoneticians measuring by hand missed synthetic F1 by 32 Hz (8%) to 118 Hz (26%) (Shadle, Nam and Whalen, JASA 2016).
- Children’s F1 discrimination threshold averaged a 49% F1 difference vs 33% for adults (Polsterer et al., JASA Express Letters 2026).
1. The Reference Datasets Behind Every Vowel Chart
Nearly every vowel chart in a phonetics textbook traces back to three American English recordings made in 1952, 1995 and 1999. Across the seven American English datasets compared in a 2018 review, mean corner-vowel formants differ by about 500 Hz for men’s F2 of /u/ and about 700 Hz for women’s F2 of /u/ (Kent and Vorperian, Journal of Communication Disorders 2018). A “norm” is only meaningful together with the dataset, dialect and decade it comes from.
The reference sets are also small by modern standards: dozens of talkers per group, not thousands. The Hillenbrand recordings and measurements are publicly available (GitHub mirror of Hillenbrand et al. 1995, hosted with the author’s permission), so anyone can rerun the classic analyses.
| Metric | Value | Source |
|---|---|---|
| Talkers and vowels in Peterson and Barney 1952 (hVd words such as heed, hid, head, had) | 76 talkers (33 men, 28 women, 15 children); 10 vowels, each said twice; 1,520 recorded vowels | Praat manual, Peterson and Barney 1952 formant table |
| Talkers in Hillenbrand et al. 1995 | 139 (45 men, 48 women, 46 children) | Hillenbrand, Journal of the International Phonetic Association 2003 |
| Steady-state vowel observations in the public h95 dataset | 1,668 observations; 10 missing F2 and 41 missing F3 values imputed | phonTools R package, h95 dataset documentation |
| Speakers in Lee, Potamianos and Narayanan 1999 | 436 children aged 5 to 17 years, plus 56 adults | Lee et al., JASA 1999 (abstract) |
| Speakers in the newest US developmental corner-vowel dataset | 190 typically developing speakers aged 4 to 20 (97 female, 93 male) | Vorperian et al., JASA 2019 |
| Spread of mean corner-vowel formants across 7 American English datasets | About 500 Hz (men’s F2 of /u/), 400 Hz (women’s F1 of /æ/), 700 Hz (women’s F2 of /u/) | Kent and Vorperian, Journal of Communication Disorders 2018 |
| Adult male vowel space area: Peterson and Barney vs 18-year-old males in Lee et al. 1999 | 411.5 kHz² vs about 330 kHz² | Kent and Vorperian, Journal of Communication Disorders 2018 |
Context note: most recent available data for the three anchor datasets: 1952, 1995 and 1999. Vowel space area means the area of the F1-F2 polygon drawn between the corner vowels; a larger area means more spread-out vowels.
2. F1, F2 and F3 Averages for Men, Women and Children
Shorter vocal tracts resonate higher, so formants climb from men to women to children. For the American English vowel /i/, mean F2 was 3220 Hz in 2-year-olds against 2146 Hz in adult men, a gap of 1,074 Hz (3220 minus 2146) (Chung et al., JASA 2012). The scaling is not uniform: the F1 of /i/ in 4-year-old boys is 1.42 times the adult male value, but its F4 is 1.57 times, so a child’s voice is not simply a man’s voice shifted up.
The Peterson and Barney rows below are the “nominal” male averages most widely reprinted; symbols and example words are given exactly as the reprint prints them. For anyone building voice effects that change perceived age or gender, the takeaway is that the formant pattern, not only pitch, differs between men, women and children.
| Metric | Value | Source |
|---|---|---|
| Peterson and Barney men, /i/ as in ‘beet’ (F1 / F2 / F3) | 270 / 2290 / 3010 Hz | Peterson and Barney, JASA 1952, as reprinted by Princeton University (COS 325) |
| Peterson and Barney men, /ae/ as in ‘bat’ (F1 / F2 / F3) | 660 / 1720 / 2410 Hz | Peterson and Barney, JASA 1952 (Princeton reprint) |
| Peterson and Barney men, /a/ as in ‘father’ (F1 / F2 / F3) | 730 / 1090 / 2440 Hz | Peterson and Barney, JASA 1952 (Princeton reprint) |
| Peterson and Barney men, /u/ as in ‘boot’ (F1 / F2 / F3) | 300 / 870 / 2240 Hz | Peterson and Barney, JASA 1952 (Princeton reprint) |
| American English /i/, adult men vs adult women (F1 / F2 / F3) | 370 / 2146 / 2838 Hz vs 391 / 2570 / 3128 Hz | Chung et al., JASA 2012 |
| American English /i/, 5-year-olds vs 2-year-olds (F1 / F2 / F3) | 439 / 2964 / 3739 Hz vs 544 / 3220 / 4210 Hz | Chung et al., JASA 2012 |
| Formant ratio, 4-year-old boys vs adult men | 1.42 (F1 of /i/); 1.57 (F4 of /i/) | Vorperian et al., JASA 2019 |
Context note: the Peterson and Barney values were read from spectrograms, while the 2012 values come from LPC analysis in Praat at the vowel midpoint, with 10 speakers (5 male, 5 female) per language and age group. Different eras, tools and dialects mean the two sets are not strictly comparable (see section 7).
3. How Identification Accuracy Depends on Formants
Formants carry most, but not all, of vowel identity. Natural vowels reach 95.4% identification, formant-synthesized copies 88.5%, and sinewave replicas only 55.0% before training (Hillenbrand et al., JASA 2011). A sinewave replica replaces each formant with a single pure tone that follows its frequency track, so it keeps the formant pattern but throws away the voice. The drop of about 40 points (95.4 minus 55.0) shows how much listeners also lean on the rest of the signal.
Talker variety costs accuracy as well. When the talkers came from two different US dialect regions and three generations, listeners fell to 80.2% overall (Jacewicz and Fox, JASA 2012).
| Metric | Value | Source |
|---|---|---|
| Natural vowels identified as intended: Peterson and Barney (10 vowel types, 76 talkers) vs Hillenbrand et al. 1995 (12 vowel types, 139 talkers) | 94.4% vs 95.4% | Hillenbrand et al., JASA 2011 |
| Formant-synthesized versions of vowels from 300 hVd syllables | 88.5% | Hillenbrand et al., JASA 2011 (citing Hillenbrand and Nearey 1999) |
| Sinewave vowels, 71 listeners before training vs after more extensive training (chance level 8.3%) | 55.0% (range 21.3% to 80.7% across listeners) vs 73.5% | Hillenbrand et al., JASA 2011 |
| Identification of 120 dialect-varied talkers (North Carolina and Wisconsin) by 30 listeners | 80.2% overall, about 15% lower than Hillenbrand et al. 1995 | Jacewicz and Fox, JASA 2012 |
| Smallest noticeable change in F1 (just noticeable difference) in 94 Dutch speakers, children aged 6-11 vs adults aged 20-36 | 49% vs 33% F1 difference (W = 1525, p = 0.001) | Polsterer et al., JASA Express Letters 2026 |
Context note: in the 2026 Dutch study the threshold was measured on each participant’s own recording of the study stimulus deur, a Dutch word kept here in the original language. Within the children, age did not significantly predict the threshold (b = -0.028, p = 0.3).
4. Age Shifts: How Children’s Formants Grow Into Adult Norms
Children are not small adults acoustically, and the sexes split earlier than voice pitch suggests. In 190 speakers aged 4 to 20, a sex difference in F2 was already present at age 4 for the high vowels /i/ and /u/ (Vorperian et al., JASA 2019, author PDF), years before puberty. The authors’ conclusion is blunt: “The acoustic properties of vowels become sex distinctive well before puberty.”
For automatic systems, child speech remains a moving target into the late teens, a challenge also visible in children’s speech recognition error rates.
| Metric | Value | Source |
|---|---|---|
| F2 sex difference: high vowels /i/ and /u/ vs the low vowels | Present at age 4 for /i/ and /u/; at ages 6 and 7 for the low vowels | Vorperian et al., JASA 2019 |
| Age at which formant data reach adult values | About 16 for female speakers, about 20 for male speakers | Vorperian et al., JASA 2019 |
| Age at which segmental durations converge to adult levels, and the lag for f0 and formant variability | Around age 12; f0 and formant variability may reach the adult range about 2 or 3 years later | Lee et al., JASA 1999 |
| Differentiation between male and female speakers, 436 children | Begins at around age 11, fully established around age 15 | Lee et al., JASA 1999 |
| Acoustic overshoot before settling at adult values | Somewhere between ages 13 and 15 | Lee et al., JASA 1999 |
| Shape of the vowel formant change with age in male speakers | Approximately linear with age | Lee et al., JASA 1999 |
| Age at which vowel production reaches adultlike precision | About 12 years | Kent and Rountrey, American Journal of Speech-Language Pathology 2020 |
| Frequency range needed to capture the first four formants of a 2-year-old | Just over 5 kHz | Kent and Rountrey, American Journal of Speech-Language Pathology 2020 |
Context note: f0 is the fundamental frequency, the rate of vocal fold vibration heard as pitch; formants are set by the vocal tract. The two develop on different schedules, which is why formant differences between boys and girls can appear long before the pitch gap. Most recent available normative child formant data: 2019.
5. Dialect and Generational Shifts
A vowel norm is a snapshot of one region at one time. Clopper and colleagues point out that the Peterson and Barney talkers “more closely approximate the speech of the American eastern seaboard than General American English” (Clopper, Pisoni and de Jong, JASA 2005), while the Hillenbrand talkers came from the northern Midwest. A formant-based classifier trained on North Carolina parents scored 88.9% on its own training set but only 44.8% on Wisconsin parents (Jacewicz and Fox, JASA 2012): moving to another dialect roughly halved its accuracy.
Change also runs across generations. In 123 women from three US regions, the total formant movement inside a vowel (its “trajectory length”, the distance the formants travel through the F1-F2 plane) was longest in the oldest group (Jacewicz, Fox and Salmons, Journal of Phonetics 2011). Dialect gaps show up in machines too: see our speech recognition accent bias roundup.
| Metric | Value | Source |
|---|---|---|
| Talkers and measurements in the six-region US dialect study | 48 talkers (4 men and 4 women from each of 6 regions); 13,440 measurements | Clopper, Pisoni and de Jong, JASA 2005 |
| Vowel by dialect interaction for F1 | F(50,2020) = 10.1, p < 0.001 | Clopper, Pisoni and de Jong, JASA 2005 |
| Listener identification of talkers from their own dialect vs the other dialect | North Carolina listeners 82.1% vs 75.8%; Wisconsin listeners 83.5% vs 79.6% | Jacewicz and Fox, JASA 2012 |
| Classifier trained on North Carolina parents: own training set vs Wisconsin parents | 88.9% vs 44.8% | Jacewicz and Fox, JASA 2012 |
| Formant trajectory length by region, 123 women | North Carolina 516 Hz, Wisconsin 396 Hz, Ohio 353 Hz | Jacewicz, Fox and Salmons, Journal of Phonetics 2011 |
| Formant trajectory length by generation: ages 51-65 vs 37-49 vs 21-33 | 477 Hz vs 408 Hz vs 380 Hz | Jacewicz, Fox and Salmons, Journal of Phonetics 2011 |
| Lowering of /æ/ across generations, North Carolina, emphatic position | F(2, 38)=29.18, p<0.001, η2=0.606 | Jacewicz, Fox and Salmons, Journal of Phonetics 2011 |
Context note: the oldest group differed significantly from both younger groups in trajectory length, while the two younger groups did not differ from each other. Recordings for the three-generation study were made from 2006 to 2008.
6. Cross-Language Vowel Spaces
Languages that share a vowel symbol do not share its formants. Adult male /u/ averaged 721 Hz on F2 in Cantonese, 1668 Hz in American English and 1704 Hz in Japanese (Chung et al., JASA 2012), so an /u/ written the same way can sit far further forward in the mouth in one language than in another. A tool that borrows English norms for another language inherits that mismatch.
Large multilingual corpora now test what stays constant. In VoxCommunis, language-average F1 values for [i] and [u] move together, a pattern called uniformity: a language that produces one high vowel with a higher F1 tends to do the same for the other (Ahn and Chodroff, VoxCommunis, LREC 2022). Coverage is still uneven, the same gap discussed in our low-resource language speech recognition roundup.
| Metric | Value | Source |
|---|---|---|
| Adult male /u/ F2 by language (5 languages, 150 speakers in total across 2-year-olds, 5-year-olds and adults) | Cantonese 721, Greek 1011, Korean 1241, English 1668, Japanese 1704 Hz | Chung et al., JASA 2012 |
| Size of the VoxCommunis corpus | 36 languages; 0.24 to 288 hours of speech and 5 to 3,620 speakers per language | Ahn and Chodroff, LREC 2022 |
| F1 correlation between [i] and [u] across languages | r = 0.83, p < 0.001, 31 languages | Ahn and Chodroff, LREC 2022 |
| F1 correlation between [e] and [o] across languages | r = 0.74, p < 0.001, 22 languages | Ahn and Chodroff, LREC 2022 |
| F2 correlation between [i] and [a]: VoxCommunis vs the earlier VoxClamantis corpus | r = 0.67 vs r = 0.06 | Ahn and Chodroff, LREC 2022 |
| UCLA Phonetics Lab Archive languages time-aligned for the VoxAngeles corpus | 95 of 314 languages | Chodroff et al., VoxAngeles, arXiv 2024 |
Context note: the VoxCommunis correlations compare language averages, not individual speakers, and the two corpora disagree on some pairs, so uniformity is a tendency, not a law.
7. Formant Measurement Error Between Tools and Methods
The numbers in sections 1 to 6 are measurements, and measurement error grows with voice pitch. In 470,000 synthesized vowels, mean LPC F1 error was 41 Hz (7%) at F0 200-300 Hz and 58 Hz (9%) above 300 Hz (Chen, Whalen and Shadle, JASA 2019). LPC (linear predictive coding) is the default formant tracker in tools such as Praat; at high pitch its estimates are pulled toward the nearest voice harmonic, which is why women’s and especially children’s formants are the least reliable.
Tool choice matters too. A comparison of Praat, WaveSurfer, TF32 and CSL found the first three accurate and comparable for synthesized vowels and adult male vowels, but less satisfactory for adult females and children (Burris et al., Journal of Speech, Language, and Hearing Research 2014).
| Metric | Value | Source |
|---|---|---|
| Typical accuracy of spectrogram and LPC formant estimation | About f0/4; about 60 Hz for a child with an f0 of 250 Hz | Kent and Vorperian 2018 |
| Absolute difference between repeated formant measurements in Hillenbrand et al. 1995 | 12 to 60 Hz | Kent and Vorperian 2018 |
| Mean LPC F1 error, synthetic vowels at F0 200-300 Hz vs above 300 Hz | 41 Hz (7%) vs 58 Hz (9%) | Chen, Whalen and Shadle, JASA 2019 |
| LPC F1 error relative to the formant value, by vowel | /u/ 23%, /i/ 18%, /æ/ 10%, /ɑ/ 8% | Chen, Whalen and Shadle, JASA 2019 |
| LPC-measured vowel space area at F0 234 Hz, as a share of the true area | 85% | Chen, Whalen and Shadle, JASA 2019 |
| Manual measurement error by four phoneticians on synthetic vowels | 32 Hz (8%) to 118 Hz (26%) | Shadle, Nam and Whalen, JASA 2016 |
| Natural-speech formant bandwidths from four analysis programs falling within ±10% of manual values | None | Burris et al., JSLHR 2014 |
| Consensus reference values remeasured after 3 months within 50 Hz; reference values abandoned as uncertain | About 80%; about 15% (mostly F3 and F4) | Derdemezis et al., American Journal of Speech-Language Pathology 2016 |
Context note: Chen and colleagues estimate that more than half of the F1 and F2 variability reported in the literature is likely due to F0 variability (simulated-to-reported ratios of 0.76 for F1 and 0.58 for F2), and expect real-speech errors to be larger than in their idealized simulations. In the Derdemezis study, Praat’s default 5500 Hz formant ceiling was lowered to 5000 Hz for adult males.
Summary: Vowel Formant Frequency Norms by the Numbers
| Metric | Value | Source |
|---|---|---|
| Talkers in Peterson and Barney 1952 | 76 (33 men, 28 women, 15 children) | Praat manual, Peterson and Barney 1952 table |
| Talkers in Hillenbrand et al. 1995 | 139 (45 men, 48 women, 46 children) | Hillenbrand, Journal of the International Phonetic Association 2003 |
| Spread of women’s mean F2 of /u/ across 7 American English datasets | About 700 Hz | Kent and Vorperian 2018 |
| Peterson and Barney men, /i/ (F1 / F2 / F3) | 270 / 2290 / 3010 Hz | Peterson and Barney, JASA 1952 (Princeton reprint) |
| American English /i/ F2: men, women, 5-year-olds, 2-year-olds | 2146, 2570, 2964, 3220 Hz | Chung et al., JASA 2012 |
| Identification of natural vowels, Peterson and Barney vs Hillenbrand | 94.4% vs 95.4% | Hillenbrand et al., JASA 2011 |
| F1 just noticeable difference, children vs adults | 49% vs 33% | Polsterer et al., JASA Express Letters 2026 |
| Age of first F2 sex difference for /i/ and /u/ | 4 years | Vorperian et al., JASA 2019 |
| Differentiation between male and female speakers, start vs fully established | Around age 11 vs around age 15 | Lee et al., JASA 1999 |
| Talkers in the six-region US dialect study | 48 | Clopper, Pisoni and de Jong, JASA 2005 |
| Cross-dialect classifier accuracy, own dialect vs other dialect | 88.9% vs 44.8% | Jacewicz and Fox, JASA 2012 |
| Formant trajectory length, oldest vs youngest generation | 477 Hz vs 380 Hz | Jacewicz, Fox and Salmons, Journal of Phonetics 2011 |
| F1 correlation of [i] and [u] across 31 languages | r = 0.83 | Ahn and Chodroff, LREC 2022 |
| UCLA archive languages time-aligned in VoxAngeles | 95 of 314 | Chodroff et al., arXiv 2024 |
| Mean LPC F1 error above 300 Hz F0 | 58 Hz (9%) | Chen, Whalen and Shadle, JASA 2019 |
| Manual F1 measurement error on synthetic vowels | 32 Hz (8%) to 118 Hz (26%) | Shadle, Nam and Whalen, JASA 2016 |
| Reference values abandoned as uncertain in a four-program comparison | About 15% | Derdemezis et al., AJSLP 2016 |
Methodology and Sources
We ran targeted searches across six angles (reference datasets, sex and age norms, identification accuracy, dialect, cross-language data, measurement error) and kept only figures we read in the primary paper, an official abstract or a named reprint. Where a page was hard to reach we read it through a text proxy but link the original.
- Peterson and Barney, JASA 1952, nominal male averages as reprinted by Princeton University (COS 325), from P. R. Cook (ed.), Music, Cognition, and Computerized Sound, MIT Press 1999.
- Praat manual: Peterson and Barney 1952 formant table (talker counts and vowel list).
- Hillenbrand, American English: Southern Michigan, Journal of the International Phonetic Association 2003.
- Hillenbrand et al. 1995 data files, GitHub mirror and phonTools h95 dataset documentation.
- Hillenbrand et al., Perception of sinewave vowels, JASA 2011.
- Kent and Vorperian, Static measurements of vowel formant frequencies and bandwidths: A review, Journal of Communication Disorders 2018.
- Lee, Potamianos and Narayanan, Acoustics of children’s speech, JASA 1999 (abstract).
- Vorperian et al., Corner vowels in males and females ages 4 to 20 years, JASA 2019.
- Kent and Rountrey, What acoustic studies tell us about vowels in developing and disordered speech, AJSLP 2020.
- Polsterer et al., Auditory acuity to vowels is more refined in adulthood compared to childhood, JASA Express Letters 2026.
- Clopper, Pisoni and de Jong, Acoustic characteristics of the vowel systems of six regional varieties of American English, JASA 2005.
- Jacewicz and Fox, Effects of cross-generational and cross-dialectal variation on vowel identification and classification, JASA 2012.
- Jacewicz, Fox and Salmons, Vowel change across three age groups of speakers in three regional varieties of American English, Journal of Phonetics 2011.
- Chung et al., Cross-linguistic studies of children’s and adults’ vowel spaces, JASA 2012.
- Ahn and Chodroff, VoxCommunis: A Corpus for Cross-linguistic Phonetic Analysis, LREC 2022.
- Chodroff et al., Phonetic Segmentation of the UCLA Phonetics Lab Archive (VoxAngeles), arXiv 2024.
- Chen, Whalen and Shadle, F0-induced formant measurement errors result in biased variabilities, JASA 2019.
- Shadle, Nam and Whalen, Comparing measurement errors for formants in synthetic and natural vowels, JASA 2016.
- Burris et al., Quantitative and descriptive comparison of four acoustic analysis systems: vowel measurements, JSLHR 2014.
- Derdemezis et al., Optimizing vowel formant measurements in four acoustic analysis systems for diverse speaker groups, AJSLP 2016.
- Data watch: the Hillenbrand et al. 1995 identification rate appears as 95.4% in Hillenbrand et al. 2011 but as 95.8% in the comparison column of Jacewicz and Fox 2012 (Table III); we cite 95.4%. Sources also disagree on timing: Vorperian et al. 2019 find formant sex differences from age 4 and f0 sex differences from age 7, the Kent and Vorperian review cites earlier data putting f0 differences at 10 or 11, and Lee et al. 1999 place the start of male-female differentiation around age 11; the studies used different ages, vowels and measures. The Peterson and Barney per-vowel values come from a reprint, not the 1952 paper itself, and only the male averages were available to us, so female and child averages from the original table are not quoted here. Most core norms in this roundup are decades old, and we could not verify a new large American English normative vowel set published in 2025 or 2026. Values from different studies used different tools and settings (see section 7), so cross-study differences of a few dozen hertz should not be read as real speech differences.
Last updated: November 8, 2026. We update this roundup quarterly, and the next refresh is expected when the next release of the VoxCommunis corpus is published.