Sarvam AI's Saaras V4 Beats Global English Benchmarks With Single-Pass Multi-Speaker Recognition

Sarvam AI used its first Epoch 2026 keynote in Bengaluru to announce Saaras V4 Multi-speaker, a speech recognition model built to handle the hardest audio scenario in real-world deployments: multiple people talking at the same time. It also marks the first time an India-built ASR system claims state-of-the-art performance on global English benchmarks.

One pass instead of three

The conventional pipeline for multi-speaker audio runs three separate steps: a diarizer figures out who is speaking when, cuts the audio into chunks, and then an ASR model transcribes each chunk sequentially. Overlapping speech gets dropped entirely. Saaras V4 Multi-speaker collapses this into a single forward pass using a speaker-biased architecture that performs diarization and transcription simultaneously, preserving overlaps and back-channels that the old pipeline would discard.

At the keynote, Sarvam demoed the model separating an overlapping male speaker and female anchor, comparing output directly against ElevenLabs Scribe v2 and Deepgram Diarization v2. The unified architecture also handles Indic languages, English, and code-mixed speech without switching between models.

English SOTA, claimed for the first time from India

Saaras V4 was evaluated on Svarah, a 9.6-hour benchmark covering 117 speakers across 65 districts in 19 Indian states, with both read and spontaneous conversational speech. The bigger claim is on global English: averaged across seven standard English benchmarks, six of them international, Saaras V4 achieves state-of-the-art performance. No India-built model has made that claim before.

The model ships with five output modes in a single architecture:

• Verbatim — word-for-word, exactly as spoken
• Transcribe — normalized output with punctuation and numerals
• Code-mixed — native script with English terms preserved in Roman
• Transliteration — full output in Roman script
• Translation — direct speech-to-English text

How it was built

Saaras V3 was trained on over one million hours of real Indian speech through a four-stage pipeline: large-scale pre-training, supervised fine-tuning, reinforcement learning, and post-training for long-tail errors. V4 extends that foundation with the speaker-biased architecture for multi-speaker scenarios, trained on one million-plus hours of curated multilingual audio spanning Indian languages, accents, and acoustic conditions, with extra attention to low-resource languages.

To support the multi-speaker work, Sarvam is open-sourcing a dataset on Hugging Face: the first multilingual Indic multi-speaker conversational corpus with heavy overlap. It covers 22 Indian languages, 108 hours of real conversational audio, and roughly 1,200 speakers. Evaluation uses two metrics: CpWER (combined diarization and ASR error rate) and DER (diarization error rate in isolation).

Where it holds up, and where it struggles

Most global ASR models are trained on clean, studio-quality audio. Real Indian audio looks nothing like that. Call center recordings arrive at 8kHz with compression artifacts. Field interviews carry traffic noise. Group discussions have overlapping speakers. Mobile calls drop out mid-sentence. Saaras V3 was trained on exactly this kind of audio, and V4 extends that robustness to multi-speaker conditions.

The harder edge cases remain: audio with many simultaneous speakers, very short utterances in noisy telephony, and languages with limited training data even within the 22-language coverage. Diarization quality degrades as speaker count rises, a known challenge across the field.

What developers can build with it

The multi-speaker capability makes a specific class of applications practical that were previously unreliable to build:

• Call center analytics — agent and customer voices separated and attributed automatically, enabling QA at scale without manual review
• Meeting transcription — panel discussions, interviews, and team calls where speakers frequently overlap
• Voice agent handoffs — detecting when a customer interrupts a bot mid-sentence, rather than waiting for silence
• Healthcare dictation — doctor-patient conversations where both parties speak simultaneously
• Media and journalism — transcribing multi-person interviews or panel recordings in Indian languages

The broader Epoch 2026 release

Saaras V4 Multi-speaker was one of several announcements at the event. Sarvam's flagship 105B model now supports voice and is priced at $0.80 per million blended tokens, roughly 5.5x cheaper than GPT-mini. Bulbul V4 text-to-speech was also announced, generating speech with emotional cues including laughter, excitement, and emphasis. The company additionally announced a sovereign inference platform, a training SDK, and smart glasses for the visually impaired.

Saaras V4 and Saaras V4 Multi-speaker are available via the Sarvam API and in the playground now. Pricing starts at Rs. 1.5 per minute, with free credits on signup. The open-source multi-speaker dataset is releasing on Hugging Face alongside the announcement.

For teams building voice products in India, the single-pass architecture removes a real integration burden. Speaker attribution previously required stitching together multiple models and accepting that overlapping speech would be lost. Getting that in one API call, across 22 Indian languages, changes what's practical to ship.