Illustration by tuput
English
Sarvam AI of Bengaluru released the Saaras V4 speech recognition model on 25 September 2026 and the Bulbul V4 Flash text-to-speech model on 13 October. Saaras V4 covers the 22 scheduled languages plus English, and Bulbul V4 Flash costs Rs 20 per 10,000 characters, a third below its predecessor. Most scores come from Sarvam itself, and AI4Bharat's public test sets sit behind several of them.
Sarvam AI, a Bengaluru company, released two voice models within three weeks of each other. Saaras V4, announced on 25 September 2026, turns speech in all 22 scheduled Indian languages, plus English, into text. Bulbul V4 Flash, announced on 13 October, turns text into speech and, according to Sarvam, plays its first sound in under 250 milliseconds at Rs 20 per 10,000 characters. Between them they cover the listening and speaking halves of a phone-based voice agent, a program that talks to a customer in Hindi, Tamil or Bengali.
Saaras V4 gives five versions of the same speech
An audio encoder feeds a language model of 3 billion parameters (the adjustable numbers inside a model) that Sarvam says it trained from scratch in-house. Its decoder is a hybrid state-space model, an alternative to the attention layers most large language models use. A step between the two shortens the audio sequence so that long recordings fit in the decoder’s working memory.
Sarvam puts the weight on the output. From one recording the model can return five kinds of text, and the launch post shows them on a single Hindi sentence about T20 batting.
| Mode | What it returns |
|---|---|
| transcribe | Native script, with numbers and dates normalised |
| codemix | Transcribe, but English words stay in English letters |
| translit | The whole sentence in English letters, which the post says is colloquially called “WhatsApp language” |
| translate | An English translation |
| verbatim | Every word as spoken, so “T20” appears as the Hindi words for “tee twenty” |
Since all five come from one model, Sarvam says no extra processing steps are needed, and so none in which errors can pile up. Developer documentation lists the same modes and names Saaras V4 as the default model. It adds recognition of global English accents alongside Indian English, and describes keyterm prompting, where a developer supplies up to 50 names or technical terms to bias recognition toward them. That feature works on the synchronous and batch interfaces but not on streaming.
Language identification error is 5.22% across all 22 languages and 2.9% across the ten most widely spoken, on verified utterances from AI4Bharat’s IndicVoices set, the post reports. For live use it gives a time to first token below 150 milliseconds when streaming. Clips under 30 seconds go through the synchronous interface, and the batch interface takes files up to two hours, with speaker labels as an option. Business Standard reported the launch on 25 September and repeated Sarvam’s figures. An earlier tuput roundup of Indian speech models covers V4’s IndicContextEval result and how Saaras V3 scored, so those are not repeated here.
Bulbul V4 Flash is built for speed and price
In its 13 October post Sarvam says Flash delivers under 250 milliseconds from request to first audio in production and runs up to 40% faster than Bulbul V3 under comparable load. The price is Rs 20 per 10,000 characters against Rs 30 for V3, and the pricing page, read on 12 November 2026, shows the same rates as Rs 2.00 and Rs 3.00 per 1,000 characters. Analytics India Magazine reported the launch on 14 October with matching numbers.
Sarvam gives two reasons for the quality. It says it expanded its own speech data nearly seven times over V3 and pretrained the model from scratch with Indian-language representation as a main goal. Its research versions, trained on more than 400 billion audio tokens (small pieces of sound, as text models use pieces of words), kept improving at that scale, and Flash is a smaller model distilled from what they taught it.
The second emphasis is reading numbers and names aloud correctly. A text normalisation layer reads each item in context and decides how it is spoken: ISRO as a word, EMI letter by letter, PAN as a word, Aadhaar in its familiar pronunciation. Demos in the post include a rail helpline call and a loan reminder that carry PNR numbers, train numbers, amounts and dates. Flash launched with 83 voices and 224 voice-and-style combinations, covering customer care, sales, collections, banking, ecommerce and education, and it runs on the Sarvam API, its Playground and Content Studio.
Sarvam’s own pages differ on the language count. The launch post and Analytics India Magazine say 11 languages. Developer documentation, read on 12 November, says the Flash voice catalogue covers Hindi, English, an English and Hindi code-mixed style, Bengali, Gujarati, Kannada, Marathi, Punjabi, Tamil and Telugu, and that Malayalam and Odia, which Bulbul V3 offers, are not yet on Flash. It also counts 222 voice IDs where the post counts 83 voices and 224 combinations. Sarvam does not explain either gap.
Why Sarvam and AI4Bharat say word error rate misjudges Indian speech
Speech scores usually use word error rate (WER), the share of words that differ from a human transcript. Sarvam’s 2 April 2026 post argues that WER, character error rate and BLEU were designed mainly for English, where each word has one accepted spelling. It gives two worked cases.
In the first, a speaker says a casual Tamil sentence meaning “they work together”, and the system writes it in the colloquial form where the reference has the formal one. Four of five words are marked wrong, a WER of 80%, though the post says a native speaker hears a perfect transcript. In the second, a Hindi sentence has “doctor” in English letters in the reference and in Devanagari in the output. One word in five is marked wrong, 20%, though the meaning is identical. A bank bot handling code-mixed Hindi, with loanwords such as account, balance and transfer, can look 15% worse than it is, the post adds.
Sarvam’s remedy is a layered set of scores. LLM-WER runs the usual comparison, then asks a language model whether each mismatched stretch means the same thing or sounds the same, and defaults to “different” when unsure. An Intent Score marks a translation 1 or 0 on whether its core meaning survived. An Entity Preservation Score counts the share of names, numbers and dates that came through, and the post recommends at least 0.95 for banking, medical or navigation uses. Gemini, through Google’s Vertex AI, is the judge, and Sarvam has published the code on GitHub. The post itself says it does not think the approach is “the final word” on Indic ASR evaluation.
Other researchers share the concern. AI4Bharat, the IIT Madras lab behind the IndicVoices and Vistaar test sets, published a paper in March 2026 for ICASSP 2026 on orthographically informed WER. Allowing valid spelling variants cut error rates by 6.3 points on average, it reports, and narrowed the gap between two systems from 18.1 to 11.5 points.
The public tests behind the numbers
Several tests named in Sarvam’s posts were built outside the company. AI4Bharat’s Vistaar, from 2023, collects 59 benchmarks across Indian languages and domains, and its authors fine-tuned Whisper models into IndicWhisper, which had the lowest WER on 39 of the 59. IndicVoices, a 2024 dataset, holds 7,348 hours of speech from 16,237 speakers in 145 districts across all 22 languages, of which 1,639 hours were transcribed when the paper appeared. For Vistaar, the Saaras V4 post reports results in ten languages as both WER and LLM-WER, and the figures appear only as charts.
On the English side, six of the seven test sets are taken as published from Hugging Face’s Open ASR Leaderboard, the post says. The seventh is AI4Bharat’s Svarah set of Indian-accented English, and normalisation and scoring follow the leaderboard’s code. That leaderboard’s paper describes a reproducible comparison of 86 systems across 12 datasets.
The keyterm figure rests on IndicContextEval, from researchers at AI4Bharat and Sarvam AI and accepted at Interspeech 2026. It has 56 hours of speech from 555 speakers in 8 languages and 23 professional domains, tested at seven levels of added context. Sarvam says Saaras V4 reaches the lowest WER, 16.03%, when a list of key terms in native script is supplied. The paper, submitted on 17 June, three months before V4 was announced, tests Sarvam Audio instead and reports 15.70% for it at that level, against 17.46% for Gemini 3 Flash and 26.04% for GPT-4o Transcribe. Those are different models, and the V4 post does not say whether its run used the same eight languages.
What has not been shown
As of 12 November 2026 tuput could read no per-dataset table for Saaras V4 in the text of Sarvam’s post, which shows sample outputs from ElevenLabs Scribe v2, Deepgram Nova3 and GPT-4o Transcribe but prints no scores for them. The claim of top results on all 22 languages is therefore Sarvam’s own. For Bulbul V3 Sarvam reported a blind listening study run by Josh Talks, which it describes as independent, with 50 to 70 annotators per language, and that study had ElevenLabs v3 alpha ahead on full-band audio quality. The V4 Flash post reports no listening score.
Sarvam’s pages also advertise voice agents from Rs 3 a minute, models included. That is a separate product from the speech models, and the pricing page lists speech to text at Rs 0.49 per minute on its pay-as-you-go plan.
Sarvam is one of the twelve teams chosen under the IndiaAI Mission, and the 13 February 2026 PIB annexure lists Rs 246.72 crore of compute support for it, with the mission’s compute and models covered in a separate piece. Neither speech post says the voice models used mission compute. Bhashini and AI4Bharat offer public speech tools for the same 22 languages, which are those in the Eighth Schedule of the Constitution, set out in our explainer on the 22 languages.
Bulbul V4 Pro, previewed by Sarvam as a larger model that can laugh, sigh and whisper, has no announced release date or price as of 12 November 2026.
Sources & further reading
- Sarvam AI: Introducing Saaras V4 (25 September 2026)
- Sarvam AI: Introducing Bulbul V4 Flash (13 October 2026)
- Sarvam AI: Evaluating Indian Language ASR (2 April 2026)
- Sarvam AI: Introducing Bulbul V3 (February 2026)
- Sarvam API docs: Saaras speech to text model page (read 12 November 2026)
- Sarvam API docs: Bulbul text to speech model page (read 12 November 2026)
- Sarvam API docs: Text to speech voices and the Bulbul v4 Flash catalogue (read 12 November 2026)
- Sarvam AI: API pricing (read 12 November 2026)
- Business Standard: Saaras V4, Sarvam AI's new speech model can transcribe 22 Indian languages (25 September 2026)
- Analytics India Magazine: Sarvam AI launches Bulbul V4 Flash for real-time speech in 11 Indian languages (14 October 2026)
- arXiv: Vistaar, Diverse Benchmarks and Training Sets for Indian Language ASR (Interspeech 2023)
- arXiv: IndicVoices, Towards building an Inclusive Multilingual Speech Dataset for Indian Languages (4 March 2024)
- arXiv: Towards Orthographically-Informed Evaluation of Speech Recognition Systems for Indian Languages (1 March 2026, ICASSP 2026)
- arXiv: IndicContextEval, a Benchmark for Evaluating Context Utilisation in Audio Large Language Models Across 8 Indic Languages (17 June 2026, Interspeech 2026)
- arXiv: Open ASR Leaderboard, Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation (revised 30 March 2026)
- GitHub: sarvamai/llm_wer, Sarvam's open-source LLM-WER evaluation code
- PIB: IndiaAI Mission progress, with Annexure I of support to foundation model builders (13 February 2026)
- AI4Bharat, IIT Madras: home page and research areas
Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.
Enjoyed this? Get the next one.
One good read at a time, straight to your inbox. No spam, unsubscribe anytime.