Sarvam's new text-to-speech model for Indian languages is 40% faster
What's the story
Sarvam AI has unveiled its latest text-to-speech (TTS) model, Bulbul V4 Flash. The cutting-edge system is optimized for real-time voice applications in Indian languages. It promises a time-to-first-audio of less than 250 milliseconds and costs ₹20 per 10,000 characters, 33% cheaper than its predecessor, Bulbul V3. The new model is also said to be up to 40% faster under similar workloads while improving pronunciation and contextual understanding.
Use cases
Ideal for customer support and banking applications
The affordability and speed of Bulbul V4 Flash make it ideal for applications like customer support, banking, payment reminders, sales, and other automated conversations.
Sarvam AI has developed a smaller model and restructured parts of its inference infrastructure to meet these needs.
The company uses a streaming-first architecture that focuses on generating the next needed speech segment instead of processing audio too far ahead.
Performance enhancements
Promises a significant performance boost
Sarvam AI claims that these changes make Bulbul V4 Flash up to 40% faster than its predecessor under similar load conditions.
However, the company hasn't shared detailed third-party benchmarks or specific concurrent-call capacity numbers in its announcement.
The actual performance in enterprise deployments will depend on the larger application and supporting infrastructure.
Language nuances
New normalization layer for better pronunciation
Sarvam AI has also introduced a real-time contextual text normalization layer with Bulbul V4 Flash.
This feature recognizes what a piece of text represents and how it should be pronounced.
The company stressed the importance of this capability, especially in financial services and customer support, where mispronouncing a transaction amount, identification number, or booking reference can lead to confusion.
Training process
Bulbul V4 Flash is built on a larger multilingual backbone
Sarvam AI's V4 research involved training larger language backbones, experimenting with audio representations, and exploring different approaches to speech generation.
The company increased its proprietary speech datasets nearly sevenfold compared to V3, covering more regional accents, speakers, emotional expressions, and conversational styles.
However, this also posed a challenge as compact multilingual models can sometimes average speech patterns across languages.
Versatility
Comes with 83 voices across 11 languages
Bulbul V4 Flash comes with 83 voices across 11 languages and 224 voice-and-style combinations.
Sarvam AI is targeting use cases from customer care and banking to e-commerce, education, collections, and narration.
The company has also previewed another model in the V4 family, Bulbul V4 Pro, which is being developed for a wider range of expressive speech capabilities.