Glossary · AI & automation
TTS and STT (text-to-speech, speech-to-text) meaning
TTS turns written text into spoken audio; STT turns speech into text. How both work inside a voice agent, why Indian speech is hard, and how to judge quality.
Where do TTS and STT sit in a voice agent?
On a phone call, STT is the ears and TTS is the mouth. The caller speaks, STT produces text, a language model decides the reply, and TTS speaks it back. Both run in streaming mode on live calls, processing audio in small pieces so the reply can start before the caller has fully finished. If either step is slow, the caller hears an awkward pause; if STT mishears, the model answers the wrong question. Every AI calling agent for sales depends on both working well in the caller's language.
Why is Indian speech harder for STT and TTS?
- Code-mixing. A single sentence may switch between Hindi and English, or Marathi and English. STT tuned for one language struggles with the switch.
- Numbers and units. "Do carton", "dedh sau", "twelve-fifty" and "₹1,250" must all resolve to the right quantity or amount.
- Names and SKUs. Product names, brand variants and town names are rarely in a general vocabulary.
- Line quality. Mobile calls from shops and highways carry traffic, crowd noise and drops.
- Natural reading. TTS has to read "₹1,250" as an Indian listener expects and pronounce Hindi words in Devanagari or Roman script correctly.
Government work on Indian-language speech technology, such as the Bhashini platform under MeitY, has widened the range of languages covered, but domain words still need testing.
A worked example: measuring STT accuracy (illustrative)
The usual measure is word error rate (WER): the number of words substituted, deleted or inserted, divided by the number of words actually spoken. Suppose a retailer says a 20-word order and the transcript gets "chhe" as "che" (a harmless spelling) and "do carton" as "doh cartoon" (two wrong words). Strict WER counts all three wrong words: 3 ÷ 20 = 15%. Only two of them change meaning, and they turn a quantity of two into nonsense, so a low overall WER can still hide a wrong order. Test the words that carry money. These figures are illustrative.
What are the common confusions about TTS and STT?
| Term | What it is | Not the same as |
|---|---|---|
| STT / ASR | Audio to text | Understanding intent, which the language model does |
| TTS | Text to audio in a synthetic voice | Voice cloning, which copies a specific person's voice |
| Streaming | Processes audio as it arrives | Batch transcription of a finished recording |
- Judging TTS by a demo sentence. Listen to it reading your actual prices, product names and addresses.
- Ignoring interruptions. Callers talk over the agent; the system must stop speaking and listen.
- Cloning a real person's voice without consent. Use licensed voices or clear written permission.
How does this show up on real sales calls?
AI Calling uses STT tuned for mixed Hindi, English and Marathi speech and a natural TTS voice, ignoring coughs and traffic so the conversation keeps moving. The guide on AI voice agents on Hindi sales calls covers what to test before going live, and why a voice agent is only as good as your price list explains why accurate data matters as much as accurate hearing.
Related terms
Read next
Where this shows up in our work. AI Calling: Takes retailer orders when reps are away and turns after-hours enquiries into booked visits. Book a 30-minute working session with an engineer.
Questions
What is the difference between TTS and STT?
TTS, text-to-speech, turns written text into spoken audio. STT, speech-to-text, turns spoken audio into written text. A voice agent uses STT to hear the caller and TTS to reply.
Is STT the same as ASR?
Yes, in everyday use. ASR stands for automatic speech recognition and describes the same technology that converts speech into text. Vendors and researchers tend to say ASR, while product teams often say STT.
Does text-to-speech work well in Hindi?
Hindi TTS has improved a lot and many voices sound natural for everyday sentences. Prices, numbers, product names and mixed Hindi-English lines still need testing with your own scripts before going live.