Getting StartedModels

Saarika

Saarika-v2.5 is our flagship speech recognition model, specifically designed for Indian languages and accents. It always transcribes the audio in the same language it was spoken. It excels in handling complex multi-speaker conversations, telephony audio, and code-mixed speech with superior accuracy across 11 languages.

Key Features

Superior Telephony Performance

Optimized for 8KHz telephony audio with enhanced noise handling and superior multi-speaker recognition capabilities.

Intelligent Entity Preservation

Preserves proper nouns and entities accurately across languages, maintaining context and meaning in transcriptions.

Automatic Language Detection

Optional automatic language identification with LID output. Use “unknown” when language is not known for automatic detection.

Speaker Diarization

Provides diarized outputs with precise timestamps for multi-speaker conversations through batch API processing.

Automatic Code Mixing

Intelligently handles mid-sentence language switches in code-mixed speech, perfect for India’s multilingual conversations.

Multi-Language Support

Comprehensive support for Indian languages with high accuracy in mixed-language environments.

Language Support

Saarika supports 11 languages with comprehensive dialect and accent coverage, including code-mixed audio support and intelligent proper noun preservation.

LanguageLanguage Code
Englishen-IN
Hindihi-IN
Bengalibn-IN
Tamilta-IN
Telugute-IN
Gujaratigu-IN
Kannadakn-IN
Malayalamml-IN
Marathimr-IN
Punjabipa-IN
Odiaod-IN

For automatic language detection, use language_code="unknown". The model will automatically identify the spoken language and return it in the response.

Performance Benchmarks

Saarika delivers exceptional accuracy across all supported languages, as measured on the VISTAAR Benchmark.

CER (Character Error Rate) Scores

Lower is better - Compared on VISTAAR Benchmark

  • Across 11 Languages: 4.96%
  • English: 4.45%
  • Hindi: 4.42%
  • 9 Other languages: 5.07%

WER (Word Error Rate) Scores

Lower is better - Compared on VISTAAR Benchmark

  • Across 11 Languages: 18.32%
  • English: 8.26%
  • Hindi: 11.81%
  • 9 Other languages: 20.15%

Detailed CER Performance by Language

CER (Character Error Rate) measures the percentage of characters that are wrong in a transcription. Lower scores are better, with 0% being perfect.

0123456CER (Character Error Rate)4.80Bengali4.45English5.92Gujarati4.42Hindi4.27Kannada5.05Malayalam4.58Marathi5.07Oria4.72Punjabi5.79Tamil5.47TeluguLanguages

Key Capabilities

Basic transcription with specified language code. Perfect for single-language content with clear audio quality.

1from sarvamai import SarvamAI
2
3client = SarvamAI(
4    api_subscription_key="YOUR_API_SUBSCRIPTION_KEY"
5)
6
7response = client.speech_to_text.transcribe(
8    file=open("audio.wav", "rb"),
9    model="saarika:v2.5",
10    language_code="hi-IN"
11)
12
13print(response)

Next Steps

Developer quickstart

Learn how to integrate the Saarika API within your application.

API Reference

Complete API documentation for speech to text endpoints.

Cookbook

Step-by-step tutorial for speech-to-text transcription.