Phoneme inventories: India and the world

A shelf of sound inventories.

A phoneme inventory is the full list of a language's phonemes: the consonants, vowels (and tones, if the language uses pitch to tell words apart) that a linguist needs to write every word unambiguously. Page 1 used minimal pairs to tell phonemes apart, and showed that Hindi has four labial stops (/p/, /pʰ/, /b/, /bʱ/); a phonological description does that for every sound of the language, and an inventory is its summary: the language's list of phonemes.

PHOIBLE 2.0 (Moran & McCloy 2019) collects such inventories from grammars and from older databases, and re-writes them in one IPA-based notation (some source conventions survive, as we will see). Its data file has one row per phoneme per inventory: 105 484 rows, 3020 inventories, 2096 ISO language codes, and 8 contributing sources:

An inventory is an analysis, not a measurement

Two linguists describing the same language can produce different lists, and both can be defensible. The commonest disagreements:

• Units or sequences? Is Hindi t̠ʃ one phoneme or /t/ + /ʃ/? Is a long pː a separate phoneme, or two /p/s? Is ai a diphthong phoneme or two vowels?
• Loanwords. Hindi /f/, /z/ and /q/ come from Persian, Arabic and English words. Some sources list them, some mark them marginal, some leave them out.
• Phonemes or allophones? If two sounds never contrast, they are one phoneme with two variants. Deciding that requires judgement, and sources differ.
• Tones. Several sources simply never record tones (more below).
• Notation. The same breathy-voiced stop appears as b̤ or bʱ; a dental stop as t̪, t̻, or a plain t.

PHOIBLE keeps every source's inventory as it is, so many languages have several: 573 languages have more than one inventory, and one language of India has 8. Any statistic therefore needs a rule for which inventory counts.

Not a random sample

Coverage is very uneven: 78% of Australian languages (thanks to one dedicated source), 52% in South America, but only 8.4% in Papunesia (New Guinea and the Pacific), where the world's linguistic diversity is densest. PHOIBLE is a convenience sample: it holds the languages someone described and someone else digitised. Related languages also share sounds by descent, so “60% of languages” partly means “a few big families”. Read every percentage below as “of the languages in PHOIBLE”, not “of all human languages”.

About 21 consonants and 9 vowels.

The typical language in PHOIBLE has 21 consonants and 9 vowels (means 23.6 and 10.3: the long right tail pulls the mean up). The medians hardly move with the inventory-choice rule (21–22 consonants, 9 vowels). English, with 24 consonants and 16 vowels in its median inventory, is a little above average on both.

Extremes are where analysis choices matter most. !Xóõ's count depends on whether a click with its accompaniment is one segment or two. Kildin Saami reaches 113 consonants in its source because palatalised and long consonants are each listed. And the largest vowel counts (up to 50) come from sources that list every long, nasalised and diphthong vowel as a unit. Treat any single “record” with care.

Tones: a lesson in missing data

In Figure 2, 23.3% of languages list at least one tone. That is an underestimate, and for a reason that has nothing to do with the languages: four sources (ea, upsid, saphon, er) never record tones (0% of their inventories do), while 69.5% of aa and 45.2% of ph inventories do. Counting only inventories from sources that record tone, 48.8% of non-Indian languages are tonal. A zero in a database can mean “absent” or “not recorded”: always check which.

/m/ is almost everywhere. Most sounds are almost nowhere.

The winners are the sounds you would guess a baby makes first: /m/ (96.3% of languages), /i/ (93.4%), /k/ (90.9%), /j/ (the “y” of yes, 89.9%), /u/, /a/, /p/, /w/ and /n/. Corner vowels (i a u) are maximally different in the vowel space, and labial, coronal and velar stops and nasals use the three most independent places of articulation (lips, tongue tip, tongue back).

Other differences in Figure 3 are worth a look: /f/ is in 43.8% of all languages but 23.6% of Indian ones (Indo-Aryan pʰ often takes its place), and /b/, /ɡ/ and /h/ are commoner in India than elsewhere.

At the other end is a very long tail. The median inventories use 2535 distinct symbols, and 1344 of them (53%) occur in only one language. Many are fine phonetic detail (an exact vowel quality, a particular release) recorded by one author; some are genuinely rare sounds, like the clicks of southern Africa or the double-articulated k͡p of West Africa.

140 languages, four big families and a few loners.

Which languages count as “languages of India”? We let Glottolog decide: a language is included if India (IN) is among the countries in its Countries field. That gives 518 Glottolog languages, of which 481 are spoken languages (the rest are sign languages and bookkeeping or unattested entries). PHOIBLE has inventories for 140 of them (29.1%), 239 inventories in all. (Matching Glottocodes directly gives 136 languages and 231 inventories; the difference is the dialect roll-up, e.g. the Kurukh, Abujmaria and Mising inventories are coded as dialects.)

The other 5 are Great Andamanese (Apucikwar), Önge (Jarawa-Onge family), Burushaski (an isolate), Northern Pashto and Uyghur. Some well-known members of India's smaller groups have no PHOIBLE inventory: the isolate Nihali, and the Tai-Kadai languages of Assam (Khamti, Ahom, Phake).

Coverage of India is a little above the Eurasian average, and it is unusually uniform: 98 of the 239 Indian inventories come from one survey, Ramaswami (1999), written by one team with one set of conventions. That is a blessing for comparing Indian languages with each other, and a reason for caution when comparing them with the rest of the world (section 7).

Retroflex, aspirated, breathy: the South Asian signature.

In 1956 Murray Emeneau called India a linguistic area: languages from unrelated families, in contact for millennia, have come to share features. The best-known is the retroflex series, made with the tongue tip curled back behind the gum ridge (ʈ ɖ ɳ ɭ ʂ ɽ, as in Hindi टमाटर ʈamaːʈar ‘tomato’). Masica (1991) adds the Indo-Aryan laryngeal system: voiceless, aspirated, voiced and breathy-voiced (murmured) stops at each place. PHOIBLE lets us put numbers on these claims.

How each property is detected

The contrasts are large and robust: under every inventory-choice rule India stays between 70% and 80% for retroflexes and 31–36% for breathy-voiced stops, against about 23% and 1% elsewhere. Counting marginal phonemes too changes the breathy figures by less than a percentage point. Nasalised vowels and vowel length are only somewhat more common in India, and the dental–alveolar contrast is rare everywhere, with one exception below.

By family: one area, different recipes

• Indo-Aryan is the home of the laryngeal system: 81% have breathy-voiced stops and 78% the full four-way contrast. Most also have retroflexes (89%) and nasalised vowels (58%). The exceptions are instructive: Assamese has merged the dental and retroflex series into alveolar t d, and the Bangladeshi-Bengali inventory shows the same.
• Dravidian is the retroflex family par excellence (100%, including the approximant ɻ of Tamil and Malayalam ழ), with vowel length in 84% and the only common dental–alveolar contrast (36%): Malayalam distinguishes n̪, n and ɳ. Aspirated and breathy stops (16%) appear in the literary languages, Telugu, Kannada and Malayalam, mainly in words borrowed from Sanskrit; Tamil has none.
• Austroasiatic: the Munda languages have retroflexes (80% of the group) and several have Indo-Aryan-style breathy stops, borrowed through long contact; the Mon-Khmer-side languages (Khasi, Pnar, Nicobarese) have at most one retroflex.
• Sino-Tibetan languages mostly have a two-way plain–aspirated contrast (87% aspirated, only 17% breathy), fewer vowels, and tone: 30% list tones, but 53% of those whose source records tone at all. Retroflexes (45%) are mostly in Himalayan languages (Ladakhi, Balti, Tibetan varieties, Kinnauri, Sherpa, Dzongkha) and rare in the Naga, Kuki-Chin and Bodo-Garo languages of the North-East.

More consonants, mostly thanks to the stops.

The median language of India has 28.5 consonants against 21 elsewhere (more than 78% of all PHOIBLE languages), and 12 vowels against 8. The consonant gap is robust (India's median ranges 27–30 under the other rules) and its source is easy to find. A four-way laryngeal contrast at five places (labial, dental, retroflex, palatal, velar) is already 20 stops and affricates, and the median Indo-Aryan inventory has exactly 20 of them, against 8 for the median language elsewhere. Indo-Aryan's median is 32 consonants; Sino-Tibetan reaches 31 with aspiration and many fricatives and affricates, while Dravidian and Austroasiatic sit at 23 and 23, nearer the world median.

Vowel counts are noisier, because they depend so much on whether long, nasalised and diphthong vowels are listed as units: the median Sino-Tibetan language of India has 8 vowels, Indo-Aryan 14 (seven or eight qualities, many also nasalised), Dravidian 13 (five qualities, short and long).

One language, several inventories.

Before looking at single languages, see how much the sources disagree about the same language:

Hindi has five inventories with 34 to 71 consonants. The biggest counts (spa and ph) list every long (geminate) consonant such as pː kʰː as a separate phoneme; the smallest (ea) lists only the core series. The ph inventory lists 11 vowels, none of them nasalised, where others list up to 23 with nasal and long vowels as units. Santali's vowel count jumps from 11 to 31 when diphthongs are listed. None of these is “wrong”; they answer the question “what are the units?” differently. This is why a single inventory per language must be chosen with a stated rule, and why the counts in section 6 are compared as medians.

Notes on the case studies

• Assamese (the language of IIT Guwahati's state) is the Indo-Aryan exception: no retroflex series, alveolar t tʰ d d̤, and a velar fricative x (from older s) that most of its neighbours lack. Its three inventories agree on 22–23 consonants.
• Hindi: the textbook four-way system at labial, dental, retroflex, palatal (affricate) and velar places, plus the retroflex flaps ɽ ɽ̤.
• Bengali: the spa, upsid and ra inventories have retroflexes; the two inventories of Dhaka (Bangladeshi Standard) Bengali, ph and ea, have alveolar t d contrasting with dental t̪ d̪ instead.
• Marathi (only ra) has dental affricates ts dz beside palatal ones, and breathy sonorants such as m̤ l̤.
• Tamil: no native voicing contrast (voiced stops appear as marginal or allophonic in ph), up to six nasals (m n̪ n ɳ ɲ ŋ in ea), and ɻ (written ʐ in ea). Malayalam adds a three-way coronal contrast. Telugu and Kannada list aspirated and breathy stops from Sanskrit loans.
• Santali and Mundari (Munda): retroflexes and borrowed breathy stops; ra marks their checked (glottalised) final stops with a creaky diacritic (p̰ k̰). One of Mundari's ra inventories is labelled Bhumij, which Glottolog treats as a Mundari dialect.
• Khasi: Austroasiatic but on the Mon-Khmer side, with aspirated stops and a single retroflex (ɭ) in its UPSID inventory; its ra inventory has none.
• Manipuri (Meitei): tones (2 in ph, 3 in ra); its breathy stops are marked marginal in ph. Bodo has a single, small UPSID inventory (15 consonants) and, like every UPSID entry, no tones listed.

Phone sets are design decisions.

A speech recogniser or synthesiser that models sounds explicitly needs a phone set: the inventory of units it predicts or generates, and a pronunciation lexicon or grapheme-to-phoneme rules that map words onto them. Everything on this page carries over:

• The inventory question is the phone-set question. Should Hindi geminates, nasal vowels, or loan sounds such as f z q be separate units? Each choice changes the number of classes to learn and the data needed for each. Sources disagree because the linguistics is genuinely open; a phone set should state its choices just as a PHOIBLE source does.
• Shared sets across Indian languages. Because the languages share so much (Figure 7), and because Brahmi-derived scripts map closely onto sounds, Indian speech projects have defined common label sets that give the same label to the same sound in different languages (e.g. the Indian TTS consortium's common label set, Baby et al. 2016). Shared units let a model trained on one language help another.
• Cross-lingual transfer. Multilingual phone recognisers such as Allosaurus (Li et al. 2020) predict universal phones and then map them onto a target language's inventory taken from PHOIBLE, so a language with little data can still get a phone recogniser. The rare-sound tail (section 3) is where such transfer struggles: a click or a breathy stop seen in only a few training languages is hard to learn.
• Code-switching with English. Indian speakers mix English words into nearly every language, and pronounce them with their own inventory: English alveolar t d are commonly realised as retroflex ʈ ɖ in Indian English, and English f or z may surface as pʰ or dʒ in some speakers. A recogniser needs a phone set, or a lexicon, that covers both languages and these mappings.
• Evaluation by sound class. Knowing which contrasts a language uses tells you where to look for errors: aspiration and breathy voice in Hindi, the three coronal nasals in Malayalam, tone in Manipuri.

End-to-end models that output characters or subwords hide the phone set, but the same distinctions still have to be learnt from the audio, and the confusions they make follow the contrasts on this page.

Five questions.