Text to Speech (Lightning)

Math notation

View as Markdown

Pass math_notation: true on a Lightning TTS request to have the normalizer read digit-flanked math operators as spoken words. Off by default because in real traffic NxN is more often a product dimension (18x24 Inch), the 24x7 idiom, or a vehicle-registration code than an actual multiplication.

Endpoint support

Available on Lightning v3.1 and v3.1_pro across every transport:

Surfacemath_notation
POST /waves/v1/tts (sync HTTP)Supported
POST /waves/v1/tts/live (HTTP SSE)Supported
wss://api.smallest.ai/waves/v1/tts/live (WebSocket)Supported

Operator lexicon

The operator word follows number_pronunciation_language, not the synthesis language. So language=en with number_pronunciation_language=hi reads 6 x 7 as “छः गुणा सात”. Localized only for Hindi and Marathi. Every other language falls back to the English words.

Glyphsen (default / fallback)himr
× x X *timesगुणागुणिले
÷ and spaced /divided byबटाभागिले
+plusप्लसअधिक
spaced - – −minusमाइनसवजा
=equalsबराबरबरोबर
^ **to the power ofकी घातची घात

Matching rules

Unambiguous glyphs fire whether glued or spaced.

  • × x X * ^ ** = +
  • Wrong-glyph x/X counts as unambiguous when a digit sits on both sides.
  • 5x3 and 5 x 3 both become “5 times 3”.
  • 2^10 becomes “2 to the power of 10”.

Ambiguous glyphs fire only when space-padded.

  • - – − and /.
  • 5 - 3 becomes “5 minus 3”.
  • 5-3 stays untouched so range and date readers still work (2-3, 2026-07-24, 020-4825-9631).
  • 6 / 2 becomes “6 divided by 2”.
  • 1/2 stays a fraction.

Operand handling. Unicode digits (including Indic digits) count as operands. The right operand may carry a currency prefix (2 x ₹99 reads correctly). An operator never bridges a line break. The pass runs after the date, phone, and version handlers, so those readings still take precedence.

Example

$curl -X POST "https://api.smallest.ai/waves/v1/tts?model=lightning_v3.1" \
>  -H "Authorization: Bearer $SMALLEST_API_KEY" \
>  -H "Content-Type: application/json" \
>  -d '{
>    "text": "The answer is 6 x 7 and the ratio is 3 - 1.",
>    "voice_id": "avery",
>    "language": "en",
>    "math_notation": true,
>    "output_format": "mp3"
>  }' \
>  --output out.mp3

With math_notation: false (default), the client hears “six ecks seven” and “three one”. With math_notation: true, the client hears “six times seven” and “three minus one”.

Protected patterns (ON == OFF)

The following patterns produce identical audio whether the flag is on or off:

  • ISO dates: 2026-07-24
  • Slash dates: 12/04/2026
  • Ranges and scores: 2-3, 2-1
  • Version strings: 1.2.3
  • Fractions: 1/2
  • Phone numbers: 020-4825-9631
  • Hindi year ranges and dates

Known limitations

Read the flag as opt-in for a reason. Enable it only if your text is genuinely math-heavy and does not carry the patterns below.

  • 24x7 idiom. Free-text 24x7 reads as “twenty four times seven”, not the “24 by 7” idiom. Channel names carrying 24x7 are handled separately via the phrase dictionary and are unaffected.
  • Product dimensions. 18x24 Inch reads as “eighteen times twenty four Inch”. There is no unit-context guard.
  • Vehicle registration and codes. GJ 27 X 2850 reads as multiplication.
  • Glued - and / never convert (by design). 5-3 stays a range and 1/2 stays a fraction.

European languages

The normalize_eu block localizes ×, ÷, and = per language: it reads × as “per”, es as “por”, fr as “fois”. The remaining wrong-glyph and extra cases (x, *, ^, **, +, -) read in English via the fallback.