Sarvam AI¶

Sarvam AI provides high-quality text-to-speech (TTS) for Indian languages, supporting 11 language codes across major regional languages. Tarash uses the sarvamai Python SDK.

TTS Only

Sarvam AI supports text-to-speech only. Speech-to-speech (STS) is not available through this provider.

Installation¶

Install tarash-gateway with the Sarvam extra:

pip install tarash-gateway[sarvam]

This installs the sarvamai SDK (>=0.1.25).

Quick Example¶

from tarash.tarash_gateway import generate_tts
from tarash.tarash_gateway.models import AudioGenerationConfig, AudioOutputFormat, TTSRequest

config = AudioGenerationConfig(
    provider="sarvam",
    model="bulbul:v3",
    api_key="YOUR_SARVAM_KEY",
)

request = TTSRequest(
    text="नमस्ते, आज का मौसम बहुत अच्छा है।",
    language_code="hi-IN",
    output_format=AudioOutputFormat(format="wav", sample_rate=24000),
)

response = generate_tts(config, request)
print(f"Request ID: {response.request_id}")
print(f"Content type: {response.content_type}")

Parameters¶

Parameter	TTSRequest Field	Required	Notes
Text	`text`	✅	The text to synthesize into speech
Language	`language_code`	✅	Must be provided. Sarvam requires a target language code (e.g. `hi-IN`). Omitting this raises a `ValidationError`.
Voice	`voice_id`	—	Maps to Sarvam's `speaker` parameter
Output format	`output_format.format`	—	Audio codec (e.g. `wav`, `mp3`)
Sample rate	`output_format.sample_rate`	—	Defaults to `24000` Hz if not specified

Extra Parameters¶

Provider-specific options can be passed through extra_params and voice_settings:

request = TTSRequest(
    text="नमस्ते दुनिया",
    language_code="hi-IN",
    voice_settings={"pace": 1.2},
    extra_params={"enable_preprocessing": True},
)

Supported Languages¶

Language	Code
Hindi	`hi-IN`
English (India)	`en-IN`
Tamil	`ta-IN`
Telugu	`te-IN`
Kannada	`kn-IN`
Malayalam	`ml-IN`
Marathi	`mr-IN`
Bengali	`bn-IN`
Gujarati	`gu-IN`
Odia	`od-IN`
Punjabi	`pa-IN`

Supported Models¶

Model ID	Notes
`bulbul:v3`	Latest version, recommended
`bulbul:v2`	Previous version

Provider-Specific Notes¶

language_code is mandatory: Unlike most other TTS providers, Sarvam requires language_code on every request. If omitted, a ValidationError is raised before the API call is made. Use the BCP-47 style codes listed above (e.g. hi-IN, en-IN).

Authentication: The api_key is passed to the Sarvam SDK as api_subscription_key. You must always provide api_key explicitly in AudioGenerationConfig.

Response format: Sarvam returns audio as a base64-encoded string. The TTSResponse.audio field contains this base64 data, and content_type is set based on the requested output format.

No streaming: Sarvam returns the complete audio in a single response. The on_progress callback is accepted but unused.