Visual Designer

Sarvam AI launches Bulbul V4 Flash, up to 40% faster, 33% cheaper. Sarvam AI launches Bulbul V4 Flash at 33% lower cost, offering up to 40% faster performance, 83 voices and support for 11 languages in real-time voice applications. 14 Oct 2026 11:26 IST Sarvam AI has launched Bulbul V4 Flash, its latest text-to-speech (TTS) model designed for real-time voice applications across Indian languages. The company says the new model delivers time-to-first-audio of under 250 milliseconds in production and costs ₹20 per 10,000 characters, compared with ₹30 for Bulbul V3. That represents a price reduction of approximately 33%. Sarvam also claims the model is up to 40% faster than V3 under comparable workloads, while improving pronunciation, speech expressiveness and contextual understanding. Faster speech at a lower cost. One of the more significant changes in Bulbul V4 Flash is the combination of lower pricing and reduced response time. At ₹20 per 10,000 characters, the model is positioned for applications such as customer support, banking, payment reminders, sales and other automated conversations where businesses generate speech repeatedly throughout the day. For these applications, even small delays can interrupt the flow of a conversation, particularly when an AI agent must retrieve information from another system before responding. Sarvam says it has developed a smaller model and redesigned parts of its inference infrastructure to address these requirements. The company uses a streaming-first architecture that prioritizes generating the next required segment of speech instead of processing audio unnecessarily far ahead. It has also optimized how the system handles simultaneous requests to improve efficiency under heavy workloads. According to Sarvam, these changes make Flash up to 40% faster than Bulbul V3 under comparable load conditions. However, the company has not disclosed detailed third-party benchmarks or specific concurrent-call capacity figures in the announcement. Actual performance in enterprise deployments will also depend on the broader application and supporting infrastructure. Why Indian speech needs more context. Beyond speed, Sarvam has focused on a problem that remains important for voice AI in India: people rarely speak in neatly separated languages. A customer service conversation may begin in Hindi, switch to English for a banking term and return to a regional expression within the same sentence. Getting the words right is only one part of generating useful speech. An AI system must also understand how those words are normally spoken. For example, ISRO is pronounced as a word, while EMI is spoken letter by letter. PAN and Aadhaar have familiar pronunciations that users expect an automated system to recognize. Similarly, railway booking references, train numbers, payment amounts and customer names require different treatment depending on the context. To address this, Sarvam has introduced a real-time contextual text normalization layer that identifies what a piece of text represents and determines how it should be spoken. The company explained the practical importance of this capability in its announcement: "An incorrectly spoken customer name, payment amount or booking reference isn't just unnatural; it can make an otherwise helpful interaction confusing." This matters particularly in financial services and customer support, where incorrectly communicating a transaction amount, identification number or booking reference can create confusion. Sarvam demonstrated the model using scenarios such as railway booking assistance, overdue EMI reminders and conversational Indian English, including exchanges that switch between languages. These demonstrations illustrate the intended applications, although they do not independently establish how the model performs across different real-world environments. What changed behind the model? Bulbul V4 Flash draws on a larger research effort focused on improving speech generation across Indian languages. Sarvam says its V4 research involved training larger language backbones, experimenting with audio representations and testing different approaches to generating speech. The company scaled its research training data to more than 400 billion audio tokens. It also expanded its proprietary speech datasets nearly sevenfold compared with V3, increasing coverage across regional accents, speakers, emotional expressions and conversational styles. But increasing the amount of training data created another challenge. According to Sarvam, compact multilingual models can sometimes average speech patterns across languages, reducing the characteristics that make each language distinct. To address this, the company pretrained V4 from scratch with Indian-language representation as a central objective. It then used a language-balanced training approach and a final voice-specialization stage to improve pronunciation, speaker identity and natural delivery. The resulting Flash model incorporates improvements identified during the larger research effort while reducing the computational demands of production deployment. The 400 billion audio-token figure relates to Sarvam's broader V4 research, rather than being presented as the training size of the smaller Flash model alone. From customer calls to regional conversations. Bulbul V4 Flash launches with 83 voices across 11 languages and 224 voice-and-style combinations. Sarvam is targeting use cases ranging from customer care and banking to ecommerce, education, collections and narration. One of the examples presented in its announcement involves a railway assistance agent handling a passenger's waiting-list inquiry, checking alternative trains and communicating booking details in a Hindi-English conversation. Another demonstrates an AI agent reminding a borrower about an overdue EMI while communicating the amount, due date and payment options. Such applications show why speech generation increasingly needs to work alongside enterprise systems rather than operate as a standalone feature. An AI agent handling a customer request may need to retrieve information, wait for a system response and then communicate the result in language the customer understands. More expressive speech ahead. Alongside Flash, Sarvam has previewed another model in the V4 family, Bulbul V4 Pro. While Flash focuses on speed, efficiency and production requirements, Pro is being developed to support a wider range of expressive speech capabilities. The company demonstrated speech patterns involving laughter, pauses, sighs, excitement, nervousness, whispering and changes in emphasis. Sarvam has not announced a commercial launch date or pricing for Bulbul V4 Pro. Meanwhile, Bulbul V4 Flash is available through the Sarvam API, Playground and Content Studio. Existing Bulbul users can move to the new model using the existing API interface. The broader significance of the launch lies in how India's voice AI market is evolving. Generating speech in regional languages is becoming only the starting point. The more difficult task is developing systems that understand how people actually speak, handle real-world information correctly and respond quickly enough for continuous conversations.

SquadStack launches Arth V2 speech model for indian-language calls. SquadStack launched Arth V2, an in-house speech-to-text model for Indian customer calls. The company reported an 8.85% semantic word error rate and 67-millisecond latency on its own benchmark, while making the model available in closed beta. 12 Oct 2026 16:35 IST SquadStack.ai has launched Arth V2, a speech-to-text model designed for Indian-language customer calls, as the company seeks to reduce transcription errors in voice-based artificial intelligence (AI) systems used for sales and customer interactions. The company said Arth V2 recorded a semantic word error rate of 8.85% and a latency of 67 milliseconds on its internal Conversational Streaming Automatic Speech Recognition (ASR) Benchmark, which uses 863 real Indian sales calls recorded over 8 kHz telephone audio. According to SquadStack, the model produced fewer meaning-changing transcription errors than several competing systems, including models from Sarvam, Google, OpenAI and ElevenLabs. These results are based on the company's own benchmark and have not been independently verified. The Noida-based company said Arth V2 has processed more than one crore minutes of live customer calls over two months. The model is deployed within SquadStack's voice AI agents for enterprise customers and is also available in closed beta for organisations seeking to test it on their own call recordings. Benchmark focuses on indian-language conversations. SquadStack said its benchmark was designed to reflect conditions in Indian contact centres, where customers frequently switch between languages, speak over agents and make calls against background noise. More than 80% of the speech minutes in the test dataset are code-mixed, while nearly 90% contain background noise, according to the company. The recordings use narrowband telephone audio, which captures a more limited frequency range than the wideband recordings commonly used in speech-recognition testing. The benchmark measures semantic word error rate, which focuses on transcription mistakes that alter the intended meaning rather than treating every word substitution as equally consequential. SquadStack reported that Arth V2 recorded fewer meaning-changing errors than Sarvam Saaras v4, Google Gemini 3.5 Transcribe, OpenAI gpt-live-transcribe and ElevenLabs Scribe v2 Realtime. However, the release also acknowledged that Soniox v5 recorded a marginally lower error rate of 8.77%, compared with Arth V2's 8.85%. The company said Arth V2 delivered its final transcript within 67 milliseconds of a customer stopping speaking on 80% of turns. It reported that the model was faster than the competing systems in its tests, although the release did not provide a complete set of comparable latency measurements for every model. SquadStack said the benchmark's test audio, reference transcripts and scoring methodology are available on Hugging Face, allowing other developers to evaluate their own speech-recognition systems against the dataset. Company builds in-house Speech Recognition. Arth V2 is the latest version of SquadStack's in-house speech-recognition technology. The company said it trained the model from scratch using more than 60 crore minutes of Indian sales conversations, covering over 10 crore unique speakers. According to SquadStack, Arth currently handles 80% of the more than 50 lakh AI calls the company makes each day. Its chief technology officer, Vikas Gulati, said the company plans to release updated versions of the model every few weeks as it identifies and addresses transcription errors in production. The approach is intended to improve the speech-recognition component of voice AI agents, where an incorrectly transcribed amount, date or customer response can affect the system's subsequent action. Apurv Agrawal, co-founder and chief executive officer of SquadStack, said transcription remains a significant failure point for voice AI agents because downstream language models depend on correctly capturing what a customer says. The company provides voice AI agents for consumer sales and said it works with more than 60 large brands, including Kotak Mahindra Bank, PhonePe, Angel One and IndiaMART. The launch comes as companies deploy conversational AI systems for customer support, sales and other telephone-based interactions. For these applications, performance on real-world telephone audio may differ from results obtained using clean recordings or general-purpose speech benchmarks. Arth V2's broader performance will depend on how it compares with competing models across independently evaluated datasets, languages and use cases beyond SquadStack's own customer calls.

Sashikumar Sreedharan to be President and CBO at Sarvam AI: Reports. Sashikumar Sreedharan has recently stepped down as Managing Director at Google Cloud India. Sashikumar Sreedharan has been roped in as President and Chief Business Officer at Sarvam AI, as per media reports. In this role, Sreedharan is expected to drive the company's business strategy and supporting its growth as it expands its AI offerings in India and global markets. Sreedharan recently stepped down from his role as Managing Director at Google Cloud India, while Amit Kumar, Managing Director of Digital Natives Business at Google Cloud India, has been appointed as interim leader. Karan Bajwa, President, Asia Pacific, Google Cloud, announced the leadership change in a LinkedIn post. Recently, Sarvam AI was in news after it raised $234 million in the first close of its $300 million Series B at a post-money valuation of $1.5 billion. HCLTech and Bessemer Venture Partners invested in the round, with continued support from existing investors Khosla Ventures and Peak XV Partners.