sarvam ai has released bulbul v4 flash, a text-to-speech model built for real-time voice agents, with a focus on indian languages, natural-sounding speech and low latency.
bulbul v4 flash key improvements
key improvements:
- faster generation: less than 250 ms to first audio, with speeds up to 40% higher than bulbul v3.
- lower cost: ₹20 per 10,000 characters, down from ₹30 — a 33% reduction.
- more voice options: 83 voices and 224 voice-style combinations across 11 languages.
- better speech quality: improved regional accents, pronunciation, emotional expression and handling of mixed-language speech.
bulbul v4 flash is designed for applications such as customer support, banking, automated calls and booking services, where response speed and accurate pronunciation are essential.
how sarvam trained bulbul v4 flash
technical improvements:
- training focused on indian languages: sarvam says the model was trained from scratch with a balanced approach to language representation, helping preserve the distinct pronunciation of each language.
- context-aware text normalization: the model adjusts how it reads names, dates, monetary amounts, abbreviations and booking references. some abbreviations are pronounced as words, while others are spelled out letter by letter.
- nearly 7× more proprietary speech data: compared with v3, the training data covers a wider range of regional accents, speakers, emotions and conversational styles.
- over 400 billion audio tokens in research experiments: sarvam tested larger models and different approaches to audio representation and generation. this figure refers to research experiments and does not necessarily reflect the training data used for the final flash model.
bulbul v4 flash aims to make voice agents faster and cheaper to run at scale. the larger speech dataset, training focused on indian languages and context-aware pronunciation could improve automated calls and customer support, particularly in multilingual settings. this is definitely good news for expanding access to text-to-speech technology across different languages.
sources
- bulbul v4 flash announcement, sarvam ai blog – latency, voices, languages, training data, text normalization
- sarvam api pricing – bulbul v4 flash at ₹20 vs bulbul v3 at ₹30 per 10,000 characters
