Skip to navigation

Saaras

View as Markdown

Saaras is our state-of-the-art speech recognition family with flexible output formats. It supports multiple output modes including transcription, translation, verbatim, transliteration, and code-mixed outputs. Saaras is built to make Indic languages LLM-comprehensible, offering accurate transcriptions and translations across 23 languages (22 Indian languages + English).

Saaras v4 is the latest and default model. It keeps the same 22 Indian languages, adds Global English alongside Indian English, and supports keyterm prompting. Always set model="saaras:v4". Saaras v3 is legacy; to move an existing v3 integration, change only the model value.

At a Glance

Model IDsaaras:v4 (default, recommended, latest)
What it doesSpeech-to-text with five output modes: transcribe, translate, verbatim, translit, codemix
Languages23 (22 Indian languages + English) with automatic language detection (full list). Also recognizes Global English alongside Indian English
APIsREST (≤30 s), Batch (≤2 hr/file), Realtime streaming (WebSocket)
Input limits30 s per REST request; WAV, MP3, AAC, FLAC, OGG and more, all limits
PricingPricing page
Best forVoice agents, call analytics, 8 kHz telephony audio, code-mixed speech
Known limitationsSee below

Keyterm Prompting is available on saaras:v4 only. Pass up to 50 names, places, brands, or technical terms as keyterms to bias recognition toward them, on REST, Batch, and both WebSocket endpoints. The examples below use it. See Keyterm Prompting.

Output Modes

Use the mode parameter to choose the output format. All five modes work on saaras:v4 and on the legacy saaras:v3. Each mode produces a different output for the same audio.

Example audio: “मेरा फोन नंबर है 9840950950”

ModeDescriptionExample Output
transcribe (default)Standard transcription in the original language with proper formatting and number normalizationमेरा फोन नंबर है 9840950950
translateTranslates speech from any supported Indic language to EnglishMy phone number is 9840950950
verbatimExact word-for-word transcription without normalization, preserving filler words and spoken numbers as-isमेरा फोन नंबर है नौ आठ चार zero नौ पांच zero नौ पांच zero
translitRomanization - Transliterates speech to Latin/Roman scriptmera phone number hai 9840950950
codemixCode-mixed text with English words in English and Indic words in native scriptमेरा phone number है 9840950950

Key Features

Keyterm Prompting

Pass up to 50 names, places, brands, or technical terms as keyterms to bias recognition toward your domain vocabulary.

Superior Telephony Performance

Optimized for 8KHz telephony audio with enhanced multi-speaker recognition capabilities.

Intelligent Entity Preservation

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

Multi-Language Support

Supports 23 languages (22 Indian + English) with optional language identification.

Speaker Diarization

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

Direct Translation

Converts speech directly to English text, eliminating the need for separate transcription and translation steps.

Global English (v4)

saaras:v4 extends English recognition beyond Indian English (en-IN) to Global English accents, on top of the same 22 Indic languages.

Language Support

Saaras v4 supports 23 languages (22 Indian languages + English) with comprehensive dialect and accent coverage, including code-mixed audio support and intelligent proper noun preservation for speech-to-English translation.

LanguageLanguage CodeLanguageLanguage Code
Hindihi-INAssameseas-IN
Bengalibn-INUrduur-IN
Kannadakn-INNepaline-IN
Malayalamml-INKonkanikok-IN
Marathimr-INKashmiriks-IN
Odiaod-INSindhisd-IN
Punjabipa-INSanskritsa-IN
Tamilta-INSantalisat-IN
Telugute-INManipurimni-IN
Englishen-INBodobrx-IN
Gujaratigu-INMaithilimai-IN
Dogridoi-IN

Language codes are optional. When not specified or set to unknown, the model will automatically detect the input language and return a language_probability score indicating detection confidence.

Additional Capabilities:

  • Includes dialects and accents of the above languages
  • Code-mixed audio support
  • Intelligent Proper Noun and Entity Preservation to ensure proper nouns, regional names, and entities are recognized and retained accurately during transcription

API Response Format

The Speech-to-Text API returns a JSON response with the following fields:

FieldTypeDescription
request_idstringUnique identifier for the API request.
transcriptstringThe transcribed text from the provided audio file.
timestampsobject or nullContains chunk-level timestamps (start_time_seconds, end_time_seconds, words, each entry is a sentence/phrase, not an individual word). Only included when with_timestamps is set to true. Word-level timestamps are not supported.
diarized_transcriptobject or nullDiarized transcript with speaker labels. Available through batch API.
language_codestring or nullBCP-47 code of the detected language (e.g., hi-IN). Returns the most predominant language if multiple are detected. Returns null if no language is detected.
language_probabilitynumber or nullFloat value (0.0 to 1.0) indicating the probability of the detected language being correct. Higher values indicate higher confidence. Returns a value when language_code is not provided or set to unknown. Returns null when a specific language_code is provided (language detection is skipped). Always present in the response.

Example Response:

{
"request_id": "20260209_abc123-def4-5678-ghij-klmnopqrstuv",
"transcript": "नमस्ते, आप कैसे हैं?",
"timestamps": null,
"diarized_transcript": null,
"language_code": "hi-IN",
"language_probability": 0.95
}

Key Capabilities

Standard transcription in the original language with proper formatting and number normalization. This is the default mode.

from sarvamai import SarvamAI
client = SarvamAI(
api_subscription_key="YOUR_SARVAM_API_KEY"
)
try:
response = client.speech_to_text.transcribe(
file=open("audio.wav", "rb"),
model="saaras:v4",
mode="transcribe", # default mode
keyterms=["Sarvam", "New Delhi", "Vistaar"]
)
print(response.transcript)
# Output: मेरा फोन नंबर है 9840950950
except Exception as e:
print(f"Error: {e}")

Limits

LimitValue
Max audio duration (real-time REST)30 seconds
Supported formatsWAV, MP3, AAC, M4A, AIFF, OGG, OPUS, FLAC, MP4, AMR, WMA, WebM (auto-detected)
Raw PCM input (pcm_s16le, pcm_l16, pcm_raw)Requires input_audio_codec; must be 16 kHz
Longer audioUse the Batch API (up to 2 hours per file)
Rate limitsSee Rate Limits

Known Limitations

LimitationDetailWorkaround
30-second cap on real-time RESTThe real-time /speech-to-text endpoint only accepts audio up to 30 seconds longUse the Batch API for longer recordings (up to 2 hours per file)
Translate mode outputs English onlymode="translate" always produces English text, regardless of input languageUse transcribe mode if you need output in the original language

Next Steps


Saaras v2.5 is deprecated. Do not use it. It receives no new features or improvements. Use Saaras v4 with mode="translate" on the /speech-to-text endpoint instead.

About Saaras v2.5

Saaras v2.5 was the previous speech translation model on the Speech-to-Text Translate endpoint (/speech-to-text-translate). It converts speech directly to English text with enhanced telephony support and intelligent entity preservation.

Key Difference: Saaras v2.5 uses the /speech-to-text-translate endpoint, while Saaras v4 uses the /speech-to-text endpoint with mode parameter support.

Key Features (v2.5)

Superior Telephony Performance

Optimized for 8KHz telephony audio with enhanced multi-speaker recognition capabilities.

Intelligent Entity Preservation

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

Multi-Language Support

Supports 11 Indian languages with optional language identification.

Speaker Diarization

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

Direct Translation

Converts speech directly to English text, eliminating the need for separate transcription and translation steps.

Translation Quality (v2.5 Benchmarks)

COMET score, a robust metric for evaluating machine speech-translations, assesses semantic accuracy, fluency, and contextual relevance. Saaras v2.5 achieves exceptional performance on the Vistaar+Indicvoices Benchmark, a dataset curated from diverse Indian language audio sources, including code-mixed content, noisy environments, and regional accents.

COMET Score Performance:

  • Across 11 Languages: 89.3%
  • English: 94.62%
  • Hindi: 91.83%
  • 9 Other languages: 88.41%

Higher is better; Compared on VISTAAR + IndicVoices Benchmark

Why COMET? It evaluates not only lexical accuracy but also how well the translation captures meaning and context, critical for Indic languages with complex structures.

Dataset Description: Contains real-world, multi-accented speech samples that covers 10 major Indic languages, ensuring representation of India’s linguistic diversity. Includes code-mixed phrases, domain-specific vocabulary, and colloquial expressions.

10095908580COMET Score88.06Bengali94.62English90.33Gujarati91.83Hindi88.3Kannada89.28Malayalam89.07Marathi89.77Odia86.39Punjabi86.45Tamil88.06TeluguLanguages