Saaras V4 puts 22 Indian languages and five transcript modes in one ASR model

Sarvam has released Saaras V4, an automatic speech recognition model for English, all 22 languages in India’s Eighth Schedule, code-mixed conversations, and noisy field recordings. A prompt selects among five output formats, including native-script transcription, romanization, and English translation.

The model uses a 3-billion-parameter hybrid state-space language decoder trained from scratch. Its architecture and prompting system let developers handle several transcription tasks through one model instead of maintaining separate services for normalization, transliteration, and translation.

Audio in, prompted text out

An audio encoder first converts speech into embeddings that capture phonetic and acoustic information. A temporal-downsampling adapter reduces the number of audio steps and projects them into the decoder’s embedding space. The Sarvam-3B decoder then receives those embeddings alongside a text prompt and generates the result one token at a time.

The hybrid state-space design processes sequences by carrying forward a compact internal state. Saaras V4 combines that sequence modeling with autoregressive generation, allowing prompts to control the output format and bias recognition toward supplied vocabulary.

Five transcript modes, one decoder

Saaras V4 performs formatting transformations inside the model. Its API exposes five output modes:

Performing these transformations during decoding can simplify downstream pipelines and reduce errors introduced by chained processing stages. The API example accepts one output_mode per request, so applications needing multiple versions of a recording may need to submit separate requests.

The benchmark case and its limits

Sarvam evaluated Saaras V4 on seven English speech-recognition benchmarks spanning Indian English, international accents, meetings, media, finance, and other recorded speech. The company reports the lowest average word error rate among the systems it compared. Word error rate, or WER, measures the insertions, deletions, and substitutions required to match a reference transcript; lower scores indicate fewer transcription errors.

The English evaluation combines datasets from Hugging Face’s Open ASR Leaderboard with AI4Bharat’s Svarah benchmark for Indian-accented English. Sarvam says it used the leaderboard’s normalization code, which matters because punctuation, casing, and number formatting can otherwise distort comparisons.

For Indian languages, Sarvam reports conventional WER alongside LLM-WER. The latter asks a language model to determine whether a mismatch changes the meaning or reflects an acceptable spelling variation. This can produce a more useful score for scripts with multiple valid spellings, although results also depend on the adjudicating model and its evaluation rules.

On the Kathbath Noisy dataset, which includes compression, clipping, and background interference, Sarvam reports an error rate below half those of Deepgram Nova-3 and GPT-4o Transcribe. Saaras V4 also performs language identification directly from audio, without a language hint. The reported error rate is 5.22% across all 22 supported Indian languages and 2.9% across the ten most widely spoken.

Rare words get a prompt

Developers can supply names, product terms, and acronyms in the relevant native script. The decoder uses that list to bias recognition toward domain-specific vocabulary, which can help with entities that appear rarely in general training data.

On AI4Bharat’s IndicContextEval benchmark, identified by Sarvam as an Interspeech 2026 evaluation, Saaras V4 records a 16.03% WER in the L5 keyword-prompting setting. That setting provides both the language and a native-script list of domain entities. Sarvam’s demonstration uses a Gujarati field recording in which the model corrects its rendering of the word for “drone” after receiving the term as a prompt.

Three routes into production

Sarvam reports streaming time to first token below 150 milliseconds. That metric measures the delay before the first partial result appears; full-transcript latency also depends on audio buffering, network conditions, and decoding time. The company also says Saaras V4 can process multi-minute recordings in under a second, though production throughput will depend on the deployment and workload.

The Sarvam API docs describe three integration paths:

• REST: synchronous transcription for recordings shorter than 30 seconds.
• Batch API: files up to two hours, with optional speaker diarization to identify who spoke when.
• WebSocket: streaming transcription with partial results for live applications.

Python and Node SDKs are available, along with integrations for LiveKit Agents, Pipecat, and the Vercel AI SDK. A Python request can select the model and output format directly while loading the API credential from an environment variable:

import os
from sarvamai import SarvamAI

client = SarvamAI(
    api_subscription_key=os.environ["SARVAM_API_KEY"]
)

with open("call.wav", "rb") as audio:
    response = client.speech_to_text.transcribe(
        file=audio,
        model="saaras:v4",
        output_mode="codemix",
    )

print(response.transcript)

What to validate

Sarvam’s benchmark results are vendor-reported, and aggregate WER can conceal large differences across languages, dialects, microphones, and acoustic conditions. A production evaluation should cover:

• Representative speakers, dialects, code-switching patterns, and background noise.
• Accuracy for names, product vocabulary, abbreviations, and other prompted terms.
• Formatting quality across the required output modes.
• End-to-end streaming latency, concurrency, and partial-result stability.
• Diarization accuracy for overlapping or rapidly alternating speakers.
• Pricing, rate limits, data retention, deployment regions, and self-hosting requirements.

Where Saaras V4 fits

Sarvam says ten of India’s 22 scheduled languages previously lacked a commercial ASR option. Saaras V4 expands that coverage through a shared model while adding code-mixed transcription, noise handling, keyterm prompting, translation, and selectable formatting.

That combination can reduce the number of components required for multilingual voice agents, contact-center analytics, field-data collection, and compliance archives. Its practical value will depend on language-specific accuracy and API performance under each application’s audio, vocabulary, and traffic patterns.