Migrating Voice Cloning from ElevenLabs to Sarvam

Last updated: September 7, 2026

This guide covers moving cloned voices from ElevenLabs Instant Voice Cloning (IVC) to Sarvam’s Voice Cloning API. For plain text-to-speech with stock voices, start from Migrating Text-to-Speech; the base URL and auth changes below are the same.

In short: re-create each cloned voice once with POST /voices/create (one clean 10-15 second clip in, a svc-{uuid} voice ID out), then swap your synthesis call to POST /voices/clone with that voice_id. ElevenLabs returns raw audio bytes; Sarvam returns JSON with base64 audio you decode. Voice settings like stability and similarity_boost have no dial here - fidelity comes from the clip and built-in quality control.

Quick start

	ElevenLabs	Sarvam
Base URL	`https://api.elevenlabs.io`	`https://api.sarvam.ai`
Auth header	`xi-api-key`	`api-subscription-key`
Auth failure status	`401`	`403` on `POST /voices/clone`; `401` on `create`/`list`/`get`/`delete`

Concept mapping

ElevenLabs	Sarvam	Notes
`POST /v1/voices/add` (IVC create)	`POST /voices/create`	Both take multipart form data and return a stored voice ID
`POST /v1/text-to-speech/{voice_id}`	`POST /voices/clone` with `voice_id`	Synthesis with a saved clone
-	`POST /voices/clone` with `ref_audio`	One-shot cloning per request; nothing stored (no ElevenLabs analog)
`files[]` (many samples)	`file` (one clip)	Pick your single best 10-15 seconds
`labels` (language, gender, accent)	`language` (required), `gender`, `style`	`language` is a required BCP-47 code, e.g. `hi-IN`
`voice_settings.speed`	`pace`
`voice_settings.stability` / `similarity_boost` / `style` / `use_speaker_boost`	-	No dials; fidelity comes from the clip plus built-in QC
`model_id`	-	The cloning model is fixed behind the API
`output_format` (`mp3_44100_128`)	`output_audio_codec` + `speech_sample_rate`	Separate codec and sample-rate fields
`remove_background_noise`	-	Clean the clip before uploading
`requires_verification`	-	No verification step; commercial use has consent requirements

Step 1: Pick one clean reference clip

ElevenLabs IVC accepts a list of samples and improves with more; Sarvam’s POST /voices/create takes exactly one file per voice. Pick your single cleanest clip:

10-15 seconds of continuous speech from the target speaker
Under 50 MB and 5-60 seconds (hard caps for POST /voices/create; the ref_audio limits for one-shot cloning via POST /voices/clone are stricter - 10 MB and 30 seconds)
Minimal background noise - there is no server-side cleanup step, so what you upload is what the clone learns from

See Prepare Reference Audio for what makes a good clip.

Step 2: Create the voice once

Before (ElevenLabs)

After (Sarvam)

curl -X POST "https://api.elevenlabs.io/v1/voices/add" \
  -H "xi-api-key: $ELEVENLABS_API_KEY" \
  -F "name=my-cloned-voice" \
  -F "files=@sample-1.mp3" \
  -F "files=@sample-2.mp3"
# {"voice_id": "21m00Tcm4TlvDq8ikWAM", "requires_verification": false}

language is required and must be a BCP-47 code from the supported languages list. The response’s auto-generated transcript is stored with the voice and reused on every synthesis call, which is why ref_text is optional when generating.

SDK equivalents: client.voices.ivc.create(...) on ElevenLabs, client.voiceCloning.createVoice(...) on Sarvam’s sarvamai SDK (Python and TypeScript).

Step 3: Generate speech with the saved voice

Before (ElevenLabs)

After (Sarvam)

curl -X POST "https://api.elevenlabs.io/v1/text-to-speech/$VOICE_ID?output_format=mp3_44100_128" \
  -H "xi-api-key: $ELEVENLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Welcome to our platform!",
    "model_id": "eleven_multilingual_v2",
    "voice_settings": {
      "stability": 0.5,
      "similarity_boost": 0.75,
      "speed": 1.0
    }
  }' --output speech.mp3

Two response-handling changes: Sarvam returns JSON, so decode response.audio from base64, and language_code is required on every synthesis call. Sample rates above 24000 Hz are neural-upsampled; see Choose Audio Formats.

SDK equivalent: client.voiceCloning.textToSpeech({ voiceId, text, language_code }) on Sarvam’s sarvamai SDK.

What Sarvam gives you that ElevenLabs doesn’t

One-shot cloning. POST /voices/clone accepts ref_audio directly - clone and synthesize in a single stateless call with nothing stored, handy for per-user or ephemeral voices.
Built-in quality control. enable_qc (on by default) transcribes the generation back with ASR, scores the character error rate against your text, and retries failing generations before returning anything.
Cross-lingual cloning in one call. Clone a voice once and synthesize across the supported languages by changing language_code - see Clone Across Languages.
Explicit language targeting. Every request carries a required language_code across 13 Indian languages, so pronunciation and prosody are correct by design rather than auto-detected.
Duration control. max_audio_duration targets a fixed output length (use it or pace, not both - sending both returns a 400).

What ElevenLabs has that Sarvam doesn’t

Professional Voice Cloning - a fine-tuned model tier trained on longer recordings; Sarvam cloning is instant-clip based.
Multiple samples per clone - Sarvam learns from one clip, so sample selection matters more.
remove_background_noise - clean your clip before uploading.
Voice sharing and library categories.
Streaming for cloned voices - Voice Cloning is synchronous REST only; for streaming, use stock-voice Text-to-Speech.

Errors and retries

Retry 429, 500, 502, 503, and 504 with exponential backoff. Do not retry 400, 401, 402, 403, 404, or 422 - those indicate problems with the request, auth, or subscription that retrying will not fix. Log the request_id from any failed response; see the FAQ.

Pricing and limits

Synthesis is billed per character of text against the Voice Cloning API; creating, listing, and deleting voices is not billed per character. See Pricing for the current rate. Rate limits for POST /voices/clone are Starter 10 / Pro 50 / Business 500 req/min - see Credits & Rate Limits.