Saaras V4 Turns One Indian Call Into Five Transcripts

Sarvam AI put Saaras V4 on Voice Agents on September 2, pairing an audio encoder with a 3-billion-parameter decoder that emits five transcript formats from one pass. The Bengaluru firm says the same clip can leave as verbatim text, normalised text, code-mixed text, romanized transliteration, or an English translation, across 22 Indian languages plus English.

That packaging is the product. Indian call software still expects Roman letters, and Saaras V4 builds that conversion into the model instead of leaving it to a second system that can add its own mistakes.

Five Formats From One Pass

A compliance archive wants every filler word. A CRM wants digits, not spoken numbers. An analytics desk wants English. A WhatsApp-style workflow wants the whole line in Roman script. Those jobs used to sit in separate tools. Saaras V4 treats them as five modes on one audio file.

THE FIVE MODES ON ONE CLIP

Sarvam calls translit “WhatsApp language.” The cricket example on the launch note makes the split obvious: transcribe keeps टी-20 in Devanagari, codemix writes T20 batting evolve in English inside a Hindi sentence, and translit renders the whole thought as “Aur jis tarah se T20 batting evolve kar rahi hai.” Translate then gives a clean English line about Virat Kohli’s season.

The same release adds multi-speaker recognition in a single pass, folding diarisation into transcription rather than running two systems in a row. That is the version meant for call recordings, meetings, and interviews, where who spoke matters as much as what was said. Sarvam showed that overlapping-speech work on July 30 with a clip of several people talking at once, the kind of messy Indian TV debate audio people actually record.

Introducing Saaras V4 Multi-speaker

Our best-in-class speech recognition model that accurately captures overlapping conversations and multi-speaker interactions, while delivering state-of-the-art performance in English across accents. pic.twitter.com/fUXgnMQiet

— Sarvam (@SarvamAI) July 30, 2026

English Benchmarks Run Through an Indian Lab

Sarvam scored Saaras V4 on seven English sets that cover meeting rooms, podcasts, studio reading, harder read speech, earnings calls, European Parliament audio, and Indian-accented English. It says the model posted the lowest average word error rate of any system it tested on that mix. Six of the sets are foreign-accent English. The seventh is Svarah, AI4Bharat’s Indian-accented English set.

THE SEVEN ENGLISH SETS

• AMI: Meeting rooms recorded on far-field microphones.
• GigaSpeech: Podcasts, audiobooks, and web video.
• LibriSpeech clean: Read speech in studio conditions.
• LibriSpeech others: Read speech with harder acoustics.
• SPGISpeech: Earnings calls packed with financial terms.
• VoxPopuli: European Parliament speeches.
• Svarah: Indian-accented English from AI4Bharat.

The six international scores follow Hugging Face’s Open ASR Leaderboard scoring code. That protocol is public, which is the right way to compare English ASR. The actual per-set error rates sit in charts on Sarvam’s note, not as a printed table, so the “lowest average” line is the company’s own summary of those runs.

On a Bengali Common Voice clip in the same note, Saaras V4 printed a clean sentence about a husband who is no longer in the room. GPT-4o Transcribe returned a broken string that does not read as Bengali. That gap is easier to see than a headline about beating GPT-4, and it is the comparison Sarvam actually published.

The best English ASR is now an Indian one: across seven standard suites, from LibriSpeech and GigaSpeech to Svarah for Indian-accented speech, it records the lowest average word error rate of any system tested.

Sarvam, Epoch announcement

22 Languages, and 10 Without a Rival Product

Saaras V4 covers 22 Indian languages plus English, with automatic language detection and output in the matching native script. On verified IndicVoices clips, language-identification error is 5.22% across all 22 and 2.9% across the ten most widely spoken. Sarvam said on September 23 that ten of those 22 languages have no commercial alternative.

The scheduled list in the API docs runs from Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Odia, Punjabi, and English through Assamese, Urdu, Nepali, Konkani, Kashmiri, Sindhi, Sanskrit, Santali, Manipuri, Bodo, Maithili, and Dogri. The second group is where commercial speech APIs usually go quiet.

LANGUAGE ID AND COVERAGE

• Language count: 23 in total, 22 Indian languages plus English, with optional BCP-47 codes such as hi-IN and ta-IN.
• Top-ten ID error: 2.9% on the ten most widely spoken Indian languages, measured on verified IndicVoices utterances.
• All-22 ID error: 5.22% when the long tail is included.
• Indic test bed: Vistaar scores on ten languages, using both raw word error rate and LLM-WER.

For Indian languages, Sarvam reports both standard word error rate and LLM-WER on the Vistaar benchmark across ten languages. Raw WER counts substitutions, deletions, and insertions. LLM-WER asks whether a mismatch changes meaning or is only a spelling or formatting difference, which is a frequent issue in Indic scripts. The two figures are meant to be read together, not swapped.

The Phone Stack That Already Depends on Saaras

This is not a lab demo looking for a first customer. Sarvam’s Circle note in July said the stack was already running more than 2 million voice conversations a day, with roughly 1 crore API calls a day on its own models, and 35 million pages digitised. At Epoch the company said its voice agents had spoken over 325 million minutes, with self-serve pricing at ₹3.50 per minute.

Equal AI is the named production case. The Android assistant answers incoming calls, figures out who is calling and why, then sends the user a summary and a recording. On August 18 Sarvam said Equal was processing around 34 million audio minutes a month on Saaras, across Indian languages, regional accents, and speakers who switch mid-sentence.

We partnered with Equal to power the speech recognition behind its AI Personal Call Assistant.

Equal AI is building a personal assistant that answers incoming calls on behalf of users. It understands who is calling and why, then shares a summary and recording after each call,… pic.twitter.com/ZwMFmdr0GQ

— Sarvam (@SarvamAI) August 18, 2026

That load is 8 kHz phone audio, traffic noise, and names that never appear in a clean read-speech set. The docs tune Saaras for exactly that: telephony, code-mixing, and entity retention. Translit exists because a lot of the software downstream was built for Roman letters. If the model can emit “mera phone number hai 9840950950” directly, a lender’s workflow does not have to learn Devanagari first.

HCLTech put $150 million into Sarvam’s June 15 Series B, the $234 million first close of a $300 million round at a $1.5 billion post-money valuation, with Bessemer Venture Partners, Khosla Ventures, and Peak XV Partners also in. The speech model is one layer in that bet. Voice agents, work agents, and an India-hosted inference layer sit around it.

Sarvam Trained the Decoder From Scratch

Saaras V4 is an audio encoder with a 3B decoder, and that decoder is a hybrid state-space language model Sarvam trained in-house. Audio goes into the encoder, a temporal-downsampling adapter shortens the sequence so long recordings fit the context window, and the 3-billion-parameter decoder then reads those audio features beside a text prompt and writes tokens one by one.

The 3-billion figure is the decoder, not a full model card. Sarvam has not published the encoder size. People who looked at the note flagged that gap, and it is a fair one: you cannot compare parameter counts with Whisper or a GPT speech model until both sides of the stack are listed. The design choice is still clear. Sarvam is spending research on the hearing stage, with a small causal decoder, rather than folding speech into one giant realtime omni model.

Prompting is how the five formats stay inside that decoder. The same audio features can be asked for verbatim, transcribe, codemix, translit, or translate without a second network. Keyterm prompting, new on v4, lets a developer pass up to 50 names, places, brands, or technical terms so the decoder is biased toward them. On September 22 that list started working on the streaming WebSocket endpoints as well as REST and batch.

Why the Docs Still Point Developers at V3

Voice Agents now route to Saaras V4 with no config change. The Speech-to-Text API is slower to move. The model page still says Saaras v3 remains the default, recommended model. Developers who want v4 have to set model=“saaras:v4”. The realtime API, as of September 3, still defaults to saaras:v3-realtime even though v4 is listed beside it.

That split is not a secret. v4’s documented English gain is Global English on top of Indian English. v3 already had the five modes and the 22 Indian languages. Keyterm prompting is v4-only. For a team whose callers are almost all en-IN and Indic, v3 is the path the docs still bless. For mixed-accent inbound, or for a glossary of product names, v4 is the one to pin.

THE SAARAS V4 ROLLOUT

1. July 30, 2026: Sarvam posts the multi-speaker Saaras V4 demo during Epoch week.
2. August 24, 2026: The technical note lands, with architecture, the five modes, and the English and Indic test plan.
3. September 2, 2026: saaras:v4 goes live on Voice Agents and is marked generally available, with automatic routing.
4. September 3, 2026: v4 reaches the Realtime API; saaras:v3-realtime stays the default there.
5. September 22, 2026: Keyterm prompting starts working on streaming WebSockets for saaras:v4.

The press wave on September 23 treated all of that as a Wednesday unveil. The model had already been in public notes for a month, and on the voice-agent path for three weeks.

The 16.03% Floor on Indian Names

The number that should travel with the SOTA line is 16.03%. On IndicContextEval, an Interspeech 2026 set, Saaras V4 posts that word error rate in the L5 keyword-prompting setting, where the domain-entity list is given in native script along with the language. Sarvam says that is the lowest WER on the set. It is still roughly one word in six on a task built around names and terms that models miss.

The Gujarati field clip on the launch note is the specimen. Without keywords the model hears ઢૂન. With the list it writes ડ્રોન, drone, which is what the speaker said. The prompt pack in that demo has 24 terms: survey, 3D model, generate, real time, kinetics, RTK, scalability, modularity, PLC, transducer, flow meters, level sensors, proximity sensor. Those are the words a field-sales or civic-survey call actually hangs on. The IndicContextEval keyword-prompting paper is where competitor scores live.

Noisy Hindi, clipped Bengali, Kannada under static, and distorted English are in the same note, each run through a different output mode. The model is built for that audio. Independent V4 numbers on the full Vistaar grid, published by someone other than Sarvam, are still scarce. Until those land, 16.03% on L5 and the in-house English average are the figures that can be checked against the notes Sarvam signed.

Frequently Asked Questions

What Audio Limits Does the Saaras V4 API Enforce?

The real-time REST speech-to-text endpoint accepts up to 30 seconds per request in WAV, MP3, AAC, AIFF, OGG, OPUS, FLAC, MP4, AMR, WMA, or WebM. Longer files go through the Batch API, which takes up to 2 hours per file. Raw PCM input must be 16 kHz and needs an input_audio_codec flag.

Which Transcript Mode Should a Developer Pick First?

Transcribe is the default and returns native-script text with numbers and dates normalised. Use verbatim when a legal or QA archive needs fillers and spoken numbers left alone, translit when the next system can only store Roman letters, codemix when English loanwords should stay in English inside an Indic sentence, and translate when the only required output is English. Translate always emits English, whatever language was spoken.

Does Saaras V4 Label Speakers in a Call?

Speaker diarisation with timestamps is available on the Batch API through a diarized_transcript field. Chunk-level timestamps can be requested on regular transcription, each chunk a sentence or phrase. Word-level timestamps are not supported, so a UI that highlights one word at a time has to be built some other way.

How Does Saaras V4 Differ From Saaras V3?

Both models share the five output modes and the same 22 Indian languages. V4 adds Global English accents on top of Indian English, and it is the only version that accepts a keyterms list of up to 50 domain words. Voice Agents now send traffic to v4 automatically. The Speech-to-Text API still recommends v3 unless the request names saaras:v4.

What Does a 16.03% Word Error Rate Mean on IndicContextEval?

L5 is the setting where the test already hands the model a native-script list of domain entities plus the language. A 16.03% word error rate means about one word in six still disagrees with the reference even after that hint. Sarvam reports it as the best score on that set; it is not a clean-read-speech number, and it is not the English average from the seven-set suite.

Can Saaras V4 Detect the Spoken Language on Its Own?

Yes. If language_code is omitted or set to unknown, the API returns a BCP-47 code such as hi-IN and a language_probability between 0.0 and 1.0. When several languages appear in one clip, the code reflects the predominant language. On verified IndicVoices audio the identification error is 2.9% on the ten most spoken Indian languages and 5.22% across all 22.

Keyterm prompting is now on the streaming sockets, Voice Agents already sit on v4, and the Speech-to-Text docs still tell a new integration to start on v3. That is the split a team has to pick, clip by clip, until Sarvam flips the default.