Sarvam AI — Indian Language Toolkit

Comprehensive AI toolkit for 23 Indian languages: TTS, STT, Document Intelligence, Translation, Transliteration, Language Detection, and Chat.

Setup

1. Get a free API key at https://dashboard.sarvam.ai
2. Set environment variable: export SARVAM_API_KEY="your-api-key"

Supported Languages

hi-IN Hindi, en-IN English, bn-IN Bengali, gu-IN Gujarati, kn-IN Kannada, ml-IN Malayalam, mr-IN Marathi, or-IN/od-IN Odia, pa-IN Punjabi, ta-IN Tamil, te-IN Telugu, ur-IN Urdu, as-IN Assamese, bodo-IN/brx-IN Bodo, doi-IN Dogri, ks-IN Kashmiri, kok-IN Konkani, mai-IN Maithili, mni-IN Manipuri, ne-IN Nepali, sa-IN Sanskrit, sat-IN Santali, sd-IN Sindhi

---

Text-to-Speech

python3 scripts/tts.py "नमस्ते, आप कैसे हैं?" --language hi-IN --speaker meera

Parameter	Default	Description
text	—	Text to convert (max 2500 chars)
--language	hi-IN	Language code
--speaker	meera	Voice name
--output	output.wav	Output file
--sample-rate	24000	Audio sample rate

Speakers — Female: Meera, Priya, Neha, Simran, Kavya, Ishita, Shreya, and more. Male: Shubh, Aditya, Rahul, Amit, Dev, Arjun, and more.

---

Speech-to-Text

Three modes: REST (quick, <30s), WebSocket (real-time streaming), Batch (long audio, diarization).

# REST — quick transcription
python3 scripts/speech_to_text.py rest audio.mp3

# WebSocket — real-time streaming
python3 scripts/speech_to_text.py websocket audio.wav

# Batch — multiple files with speaker diarization
python3 scripts/speech_to_text.py batch audio1.mp3 audio2.mp3 --diarization --num-speakers 3 --output-dir ./transcripts/

Batch workflow: create job → upload files → start → poll status (Accepted → Pending → Running → Completed) → download results.

Formats: WAV, MP3, AAC, AIFF, OGG, OPUS, FLAC, MP4/M4A, AMR, WMA, WebM, PCM

---

Document Intelligence

Extract text from PDFs and images (JPEG/PNG).

python3 scripts/document_intelligence.py document.pdf --language hi-IN --format md
python3 scripts/document_intelligence.py --job-id <id> --download -o ./output/

Formats: md (default), html, json. Max 200 MB, 500 pages.

---

Translation

# Auto-detect source, translate to Hindi
python3 scripts/text_processing.py translate "Hello, how are you?" --target hi-IN

# Mayura model with colloquial mode
python3 scripts/text_processing.py translate "What's up?" --target hi-IN --model mayura:v1 --mode modern-colloquial

Models: sarvam-translate:v1 (23 languages), mayura:v1 (12 languages, supports modes and transliteration)

Modes (mayura only): formal, modern-colloquial, classic-colloquial, code-mixed

---

Transliteration

python3 scripts/text_processing.py transliterate "नमस्ते" --source hi-IN --target en-IN
python3 scripts/text_processing.py transliterate "namaste" --source en-IN --target hi-IN --spoken-form

---

Language Detection

python3 scripts/text_processing.py detect "नमस्ते दुনিয়া"
# Output: Language: hi-IN, Script: Deva

---

Chat Completion

Two models: sarvam-105b (flagship, complex reasoning) and sarvam-m (efficient, general chat).

python3 scripts/text_processing.py chat "Explain quantum computing" --model sarvam-105b
python3 scripts/text_processing.py chat "What is the capital of India?" --model sarvam-m --temperature 0.8

---

Resources

• Dashboard: https://dashboard.sarvam.ai
• Docs: https://docs.sarvam.ai
• Cookbook: https://github.com/sarvamai/sarvam-ai-cookbook