Voice APIs for companies building in many languages: speech‑to‑text, text‑to‑speech and voice‑to‑voice models that understand code‑switching. Speech‑to‑text is live today in 22 Indian languages and English.
Speech-to-text API22 Indian languages + English
We train our own models and offer them to businesses through an API. Speech to text is available today; more models are on the way.
Speech to text
Turns recorded speech into accurate text, including speech that switches languages mid-sentence.
Text to speech
Turns written text into natural speech, including mixed-language text such as Hinglish, read the way a person would say it.
Voice to voice
A single model that listens and replies in speech, with no text step in between — for natural conversations in Indian languages.
Multimodal
Models that understand speech, text and images together — and can act on what they understand, in the real world.
People across India switch languages in the same sentence. Our models are trained for it, and write each word in the script people would use.
Scriba transcribes all 22 languages in the Eighth Schedule of the Constitution of India, plus English — each in its own script.
Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam, Odia, Punjabi, Assamese, Maithili, Konkani, Nepali, Sanskrit, Santali, Dogri, Manipuri, Bodo, Kashmiri, Sindhi and Urdu.
Drawing based on the wheel of the Konark Sun Temple, 13th century
The same speech-to-text model, set up for how each team works.
Every call transcribed, with agent and customer kept apart, ready for quality review.
Speaker-labelled notes, in the language people actually spoke.
Timed transcripts, ready to turn into captions.
A word-for-word record, with nothing smoothed over.
Customer calls written correctly, with amounts and dates as digits.
Callers understood in their own language and script.
नमस्ते, बताइए मैं आपकी क्या मदद कर सकती हूँ?
मेरा card block हो गया है
कोई बात नहीं, आपका registered mobile number बताइए
जी, 98201 45673
इस week का target 120 calls है
Pune team को दो लोग और चाहिए
ठीक है, Friday तक hiring plan भेजती हूँ
আজকের ক্লাসে আমরা জলবায়ু পরিবর্তন নিয়ে কথা বলব
প্রথমে উনিশশো নব্বই সালের তথ্য দেখি
তারপর আমরা বর্তমান পরিস্থিতি নিয়ে আলোচনা করব
अं, साक्षीदारांनी, साक्षीदारांनी त्या दिवशी काय पाहिलं ते सांगावं
हो, मी संध्याकाळी सहा वाजता तिथे होतो
என் EMI ₹4,500 இந்த மாசம் கட்டலை
பரவாயில்லை, 15 November வரைக்கும் time இருக்கு
ମୋର ରାସନ କାର୍ଡ ଏପର୍ଯ୍ୟନ୍ତ ଆସିନାହିଁ
ଆପଣଙ୍କ ଆବେଦନ ନମ୍ବର କହନ୍ତୁ
The research problems behind our models.
One model that listens while it speaks — so people can interrupt, pause and talk over it, as in a real conversation.
Recognising speech that switches languages mid-sentence, and writing it in native script, mixed script or word for word.
कल की meeting का time बदलकर 4 बजे कर दीजिए
Each language in its own script, including names, numbers and places.
Hours of audio in one go, with speakers labelled consistently throughout.
Who spoke when, and each channel of a phone call transcribed separately.
Public test sets and the standard word error rate (WER), reported for every language.
Voices that read Indian languages and mixed-language text the way a person would say it.
Text that appears while people are still speaking, for live calls and captions.
How we measure. Every language on public test sets the model never trained on, with the standard word error rate. Results are in our model card — request it.
Scriba transcribes audio in 22 Indian languages and English through one API.
