Sarvam AI 发布 Bulbul V4 Flash：11 种印度语言实时语音，首音频延迟低于 250 毫秒

Sarvam AI has launched Bulbul V4 Flash, a text-to-speech model that generates speech in 11 languages with a time-to-first-audio latency of under 250 milliseconds.

The model is priced at ₹20 per 10,000 characters, a one-third reduction from the ₹30 charged for Bulbul V3.

The Bengaluru-based AI company said the new model improves pronunciation, pacing and contextual delivery while reducing the time and cost required to generate speech for real-time applications, including customer service and voice agents.

“With V4 Flash, our research focused on capturing the richness and nuances of Indian languages, while making high-quality speech generation faster and more affordable,” the company said in its announcement.

Bulbul V4 Flash is available through the Sarvam API, Playground and Content Studio. It offers 83 voices and 224 voice-and-style combinations for applications across banking, ecommerce, education, sales, customer support and narration.

The model is up to 40% faster than Bulbul V3 under comparable workloads, according to Sarvam. The company has also optimised its inference infrastructure to handle multiple simultaneous conversations.

Existing Bulbul users can migrate to V4 Flash through the existing Bulbul API.

Sarvam said it expanded its proprietary speech datasets nearly sevenfold compared with V3, increasing coverage across languages, regional accents, speakers, emotional registers and conversational styles.

The company pretrained V4 from scratch with a focus on Indian language representation and used a language-balanced training process to improve pronunciation and preserve the characteristics of individual languages.

“Regional voices sound native, Indian English sounds Indian, and names, numbers and everyday expressions come out the way people say them,” the company said.

The model also uses a contextual text normalisation layer to determine how names, dates, amounts and abbreviations should be spoken. For instance, it distinguishes between abbreviations that are pronounced as words and those spoken letter by letter.

This is particularly relevant for voice agents handling customer information, where mispronouncing a name, payment amount, or booking reference can cause confusion.

Sarvam said Bulbul V4 Flash draws on research involving larger language backbones, different audio representations and generative approaches, with training data scaled to more than 400 billion audio tokens.

The company used the results of these experiments to develop a smaller model that retains key speech-quality improvements while meeting production requirements for latency and cost.

Its streaming-first inference system prioritises audio that needs to be generated next rather than processing speech far ahead of the conversation. Sarvam said it also reduced unnecessary synchronisation and data copying to improve performance when handling concurrent requests.

Sarvam also previewed Bulbul V4 Pro, another model in the V4 family that is being developed to support more expressive speech.

The preview demonstrated the ability to convey emotions such as happiness, sadness, fear and anger, alongside actions such as laughing, sighing, whispering and emphasising specific words.

Sarvam has not announced a release date or pricing for V4 Pro.

The preview signals the company's work towards voice models that can adjust not only pronunciation and pacing but also the emotional delivery of speech, particularly for conversational AI applications.