Hindi Text to Speech: Devanagari, Hinglish and Voices
Hindi text to speech guide: Devanagari vs Hinglish input, schwa deletion, lakh and crore numbers, and voice coverage across major vendors as of October 2026.
Hindi text to speech works with Devanagari, romanised Hindi and mixed Hinglish, but results depend on the tool and how you type. Hindi words carry an inherent vowel that is often not pronounced. As of October 2026, Azure lists one Hindi locale with 17 voices, and Google offers 30 Chirp 3: HD voices.
Devanagari, romanised Hindi or Hinglish: which should you type?
The script you type changes what the voice has to decode. The table compares the three main options.
| Input | Example | What the sources say | |---|---|---| | Devanagari | नमस्ते दुनिया | Standard script, a left-to-right abugida in the Unicode block U+0900 to U+097F. Polly's bilingual voices read it | | Romanised Hindi | namaste duniya | Polly's documentation says its bilingual voices read romanised Hindi, which it calls Romanagari | | Hinglish (mixed) | "Meeting kal hai, please confirm karo" | Polly says its bilingual voices read mixtures of Hindi and English. A 2021 study of YouTube comments found 52% romanised Hindi, 46% English and 1% Devanagari Hindi |
The comment split is a secondary source covering one platform, so do not treat it as how all Hindi speakers type. Type the way your audience writes, then test the result. Hinglish is common in everyday writing, which is why the next question matters: how the voice reads it.
Why is Hindi hard for text to speech?
Two features cause most problems: schwa deletion and nukta letters.
Schwa deletion. Each Devanagari consonant carries an inherent schwa, which is often not pronounced. The word नमक is pronounced namak, not namaka, and राम is Ram. The script does not mark these deletions, so a TTS system needs deletion rules or a model that learns them. A classic rule-based deletion rule is reported to predict correctly about 89% of the time, according to Wikipedia's article on schwa deletion. That figure describes that classic rule, not modern TTS accuracy.
Nukta letters. Devanagari has letters with a dot, called nukta, for sounds such as q (क़), kh (ख़), gh (ग़), z (ज़), r-dot (ड़), rh-dot (ढ़) and f (फ़). When a word needs one of these sounds, type the nukta letter. The plain letter gives a different sound.
How are lakh and crore numbers read?
Indian numbering groups digits differently from the Western system. The first group has three digits, then the digits group in pairs. Wikipedia's article on the Indian numbering system gives these values.
| Value | Western grouping | Indian grouping | Hindi name | |---|---|---|---| | One hundred thousand | 100,000 | 1,00,000 | lakh (लाख) | | Ten million | 10,000,000 | 1,00,00,000 | crore, Hindi karor (करोड़) |
Amazon Polly's documentation says Arabic numerals are read in Indian English by default. To have them read in Hindi, you must use the hi-IN language code. Devanagari numerals are expanded in Hindi. Write the numbers the way you want them spoken, and test a large figure such as 1,00,000 before you publish.
Which tools offer Hindi voices?
The table shows what each vendor lists for Hindi, as of October 2026.
| Vendor | Hindi coverage | Notes | |---|---|---| | Microsoft Azure AI Speech | One locale, hi-IN, with 17 voices: 9 Standard, 4 MAI-Voice-2 and 4 MAI-Voice-2-Flash | The MAI-Voice-2 voices are in public preview | | Google Cloud Text-to-Speech | hi-IN with 30 Chirp 3: HD voices (14 female, 16 male), plus 16 older voices | Chirp 3: HD supports custom pronunciation with IPA or X-SAMPA for hi-IN | | Amazon Polly | Aditi (female, Standard) and Kajal (female, Neural) | Documented as fully bilingual in Indian English and Hindi | | ElevenLabs | Hindi on Multilingual v2 and Flash v2.5, and in the v3 and v4 language lists | Per-model coverage comes from the ElevenLabs models page | | Gemini API text-to-speech | Hindi listed by name; input language detected automatically | 30 prebuilt voices |
Only Polly documents Hindi-English code-switching by name in the pages reviewed. Azure's multilingual voices and Gemini's automatic language detection suggest cross-language reading, but those pages do not discuss Hinglish. Treat that as a reason to test, not as a guarantee.
Script tips for Hindi text to speech
- Type the way your audience writes. Test a 20-second sample in Devanagari and in your Hinglish version, and keep the one that reads better.
- Use punctuation for phrasing. Full stops and commas control pauses.
- Respell names that come out wrong, then listen again.
- Write numbers in Indian grouping when the audience expects it, and check large figures.
- Check the output by transcribing it. OndaLabs's speech to text is listed at 50 credits per minute, which makes a quick check affordable. Use the transcript to spot words the voice misread.
Hindi text to speech with OndaLabs
OndaLabs has a dedicated Hindi text to speech page. Its 30 prebuilt voices are usable in Hindi, delivery is directed in plain words with supported expressive tags, and the dialogue tool handles two-speaker scripts. Custom voices are available from the Creator plan at $22 per month, with 3 custom voices. They can be cloned from a recording you have permission to use or designed from a text description.
The Free plan gives 2,000 credits per month, about 2 minutes on Onda Lite at 1 credit per character. Starter at $9 adds Onda Flash, the most expressive model at 1.5 credits per character, plus commercial use. The verified feature list does not cover Hinglish handling or lakh and crore settings, so test those yourself. You can check a speech to text transcript of any output, and compare plans and prices before you choose.
For related reading, see the Arabic text to speech guide and the Japanese text to speech guide.
Frequently asked questions
Can text to speech read Hindi in Devanagari?
Yes. Several vendors list Hindi, and Amazon Polly's documentation says its bilingual voices read Devanagari, romanised Hindi and mixtures with English. As of October 2026, Azure has a Hindi locale with 17 voices, and Google has 30 Chirp 3: HD voices for hi-IN. Listen for schwa deletion in your sample, since the script does not mark it.
Can TTS read Hinglish?
Only Polly documents this by name in the sources reviewed, through its bilingual Hindi and Indian English voices. The other vendor pages do not discuss Hinglish. Test your own Hinglish script, because a bilingual voice can still mispronounce a given word.
Which AI voice sounds most natural in Hindi?
The sources reviewed do not rank Hindi voices for naturalness, and no independent Hindi benchmark was included. Compare candidates on your own script, listening for schwa deletion, numbers and names. The vendors offer different counts, with 17 voices at Azure, 30 Chirp 3: HD voices at Google, and 2 voices at Polly.
How are lakh and crore read by text to speech?
Indian numbers group digits differently. 1,00,000 is one lakh, and 1,00,00,000 is one crore. Polly's documentation says Arabic numerals are read in Indian English by default and need the hi-IN language code to be read in Hindi. Test large numbers before you publish.
Is there a free Hindi text to speech?
OndaLabs has a Free plan with 2,000 credits per month, about 2 minutes on Onda Lite, plus expressive tags and dialogue. Google lists a free tier of 0 to 4 million characters per month for WaveNet and Standard voices, and Amazon Polly lists 5 million characters per month for its Standard voices. Free tiers change, so check each pricing page.