Sarvam AI
Sarvam Motif

Speech to Text, perfected for Hindi

Understands Hindi even with background noise, multiple accents, or mid-sentence language switches.

Speak Hindi to see live captions

Every Hindi word, captured accurately

Code-mixing. Numbers. Proper nouns. Abbreviations.
0:00
हाँमैंनेPROPER NOUNFlipkartसेNUMBER₹12,499काENTITYSamsungGalaxyCODE SWITCHफ़ोनऑर्डरकियाथा।ऑर्डरIDहैABBREVIATIONFLK-78234-XN।CODE SWITCHडिलीवरीएड्रेसहै42,PROPER NOUNKoramangala5thBlock,PROPER NOUNBangalore।मेराफ़ोननंबरहैPHONE NUMBER9840950950।CODE SWITCHप्लीज़डिलीवरीजल्दीकरदीजिए,मेरापुरानाफ़ोनबिल्कुलकामनहींकररहा।

Flexible output. Same Hindi audio,
multiple formats

Transcribe

With formatting and number normalization.

Output
मेरा फोन नंबर है 9840950950

Translate

From Indic languages to English.

Output
My phone number is 9840950950

Transliteration

Indian languages written in English letters.

Output
mera phone number hai 9840950950

Verbatim

Preserves fillers and spoken numbers.

Output
मेरा फोन नंबर है नौ आठ चार zero नौ पांच zero नौ पांच zero

Saaras hears हिन्दी the way you speak it

Every feature shown with real Hindi audio. Play each clip. Read the transcript. See the difference.

Speaker Diarization

Know who said what

Automatically labels speakers in multi-party Hindi conversations. Colors match the transcript.

Click to play
0:00Speaker 1

नमस्कार डब्ल्यूसी बैंक में संपर्क करने के लिए आपका धन्यवाद। मैं गीतिका आपकी किस प्रकार सहायता कर सकती हूँ?

0:09Speaker 2

नमस्कार मैं एक गिरवी लोन के आवेदन पूरा करना चाहती हूँ कृपया मुझे प्रक्रिया बताइए

0:17Speaker 1

ज़रूर कृपया अपना अकाउंट नंबर और पैन कार्ड नंबर बताइए

Code-Mixing

Seamless Hinglish

Handles mid-sentence switching between Hindi and English. No drops, no accent shifts.

Click to play
आज का Hinglishमीटिंग बहुत Code-switchप्रोडक्टिव रहा। लंच के बाद Englishकोडिंग सेशन में सब कुछ स्मूथ चला। टीम ने बहुत अच्छा काम किया।
Telephony-Optimized

Built for real calls

8kHz call center audio, compression, background noise — maintains accuracy for Hindi where others fail.

8kHzNoisy
Click to play

अच्छा, मौजूदा सरकार का कामकाज आपको कैसा लगता है? मोदी सरकार बढ़िया है। विधायक का कामकाज कैसा रहा?

Streaming API

Real-time transcription

WebSocket streaming with sub-250ms latency. Words appear as they're spoken in Hindi.

Click to play
~90ms

भारत की सुबह का नज़ारा ही कुछ और होता है। चाय की चुस्की के साथ अखबार पढ़ना, बच्चों की किलकारियां सुनना।

Developer-first platform

OpenAI-compatible APIs. Drop-in SDKs for Python and Node.js. Go from zero to first transcription in under 5 minutes.

REST & WebSocket APIs

Standard REST for batch, WebSocket for real-time streaming with sub-250ms first byte.

SDKs & libraries

Official Python and Node.js SDKs with TypeScript support. pip install sarvam-ai.

Complete documentation

Interactive API reference, code samples, and integration guides for every endpoint.

Free tier included

Start building immediately. No credit card, no sales call, no minimum commitment. 

from sarvamai import SarvamAI

client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")

# Transcribe audio
response = client.speech_to_text.transcribe(
    file_path="call_recording.wav",
    language_code="hi-IN",
    model="saaras:v3",
    with_diarization=True
)

print(response.transcript)
for turn in response.turns:
    print(f"[{turn.speaker}] {turn.text}")

Enterprise-ready. Responsible AI.

Built with safety, compliance, and data sovereignty at the core.

SOC 2 Type II & ISO 27001

Enterprise-grade security certifications. Annual audits, documented controls, continuous monitoring.

No training on your data

Your API inputs are never used for model training. Zero data retention after processing unless explicitly requested.

Data sovereignty

All data processed and stored in India. No cross-border transfers. Full compliance with Indian data regulations.

Audio data privacy

Audio data is processed in real-time and never stored unless explicitly requested. Complete privacy by default.

Content safety filters

Automated detection and filtering of harmful, abusive, or misleading content in transcriptions.

Audit-ready logging

Comprehensive API usage logs, access controls, and RBAC for enterprise governance and compliance reporting.

Simple, transparent pricing

Start free. Scale as you grow. No hidden costs.

Base plan

₹1.25 per minute

Free trial included

No credit card required. Get API keys instantly.

Volume discounts available
Enterprise pricing available
Speaker diarization included
Usage analytics
All 22 languages
Best for startups

Your questions, answered
Saaras v3 achieves industry-leading word error rates for Hindi, including conversational speech, regional accents, and noisy environments. It handles Hinglish code-mixing natively without accuracy degradation.
Yes. Saaras v3 is trained on diverse Hindi speech data covering major dialects and regional accents from across India, including Bhojpuri-influenced Hindi, Rajasthani Hindi, and more.
Absolutely. Our WebSocket streaming API delivers sub-250ms latency for Hindi, making it ideal for real-time voice agents, IVR systems, and conversational AI.
Saaras v3 was trained on natural Hindi-English code-switching patterns. It transcribes mid-sentence language switches accurately without dropping words or misattributing languages.
10+ formats including MP3, WAV, AAC, OGG, Opus, FLAC, M4A, AMR, WMA, and WebM. Streaming API supports WAV and raw PCM at 16kHz.

Transcribe in 22 Indian languages Powered by Saaras v3