Speech to Text for Indian Languages

Saaras v3 automatically detects the spoken language without requiring you to specify it upfront. When a speaker switches between Hindi and English mid-sentence, or drops Tamil words into an otherwise English conversation, the model follows along without losing accuracy at language boundaries. This is not a pipeline that detects language segments and routes them to separate engines. It is a single unified model that understands multilingual Indian speech as a continuous stream.

An estimated 350 million Indians speak English as a second language, and the vast majority mix it freely with their primary language. A customer might say "Mera order ka status kya hai, I placed it last Tuesday." A doctor might dictate "Patient ko high fever hai, CBC and LFT karwao." This is normal Indian communication, and Saaras v3 transcribes it accurately because it was trained on exactly this kind of speech.

Most global ASR models are trained on clean, studio-quality audio and degrade significantly on real-world recordings. Indian production audio is rarely clean: call center recordings at 8kHz with compression artifacts, field interviews with traffic noise, group discussions with overlapping speakers, and mobile phone calls with network dropouts. Saaras v3 is trained on this exact kind of audio, so it maintains accuracy where other models break down.

The model handles ambient noise, crosstalk, varying microphone quality, and poor network connections without requiring pre-processing or noise cancellation on your end. This robustness comes from a 4-stage training pipeline: large-scale pre-training on diverse acoustic conditions, supervised fine-tuning, reinforcement learning, and a post-training stage specifically targeting long-tail errors and output stability. For voice agent deployments where callers are speaking from busy streets, factory floors, or crowded homes, this robustness is not a nice-to-have. It is what determines whether the system works in production or only in demos.

Speaker diarization answers the question "who said what?" in a multi-speaker recording. Sarvam's Batch API identifies different speakers in your audio and tags each word with a stable speaker ID. This is essential for meeting transcriptions where you need to attribute action items, call center analytics where you need to separate agent and customer utterances, and interview recordings where you need to distinguish between interviewer and interviewee.

Diarization works across all 22 supported languages and handles overlapping speech, speaker turns of varying length, and conversations where speakers have similar voice characteristics. The speaker labels are consistent across the entire recording, so Speaker 1 at minute 2 is the same person as Speaker 1 at minute 45.

Voice dictation — speaking instead of typing — is one of the most natural ways to interact with digital devices. Sarvam's speech recognition engine powers voice typing in Indian languages, letting users dictate messages, emails, notes, and documents in Hindi, Tamil, Telugu, or any of the 22 supported languages. Unlike generic voice dictation tools that force you into English, Sarvam handles dictation in Hindi with correct Devanagari output, adds smart punctuation automatically, and keeps up with natural speaking speed. For users who type faster by talking, or who find keyboards difficult to use, voice-to-text dictation removes the friction between thought and written word.

You can convert speech to text online using the playground above, or integrate via API for production use. Sign up at sarvam.ai/try/stt-api to get your API key. Install the Python SDK with pip install sarvamai or the Node.js SDK with npm install sarvamai. Your first audio to text conversion takes under 5 minutes from signup to working output.