Alibaba's Qwen-Audio 3.1 Slashes Voice AI API Prices by 95%

Building a voice bot that understands more than one Indian language has always come with a catch: the moment you scale it past a demo, the API bill catches up with you. Alibaba just cut that bill by as much as 95 percent.

At its Apsara Conference in Hangzhou last week, Alibaba's Qwen team unveiled Qwen-Audio-3.1, a family of five speech and audio models, alongside steep price cuts across the board: speech recognition (the tech that turns spoken words into text, commonly called ASR) drops by up to 95 percent, real-time voice conversation APIs by around 85 percent, and text-to-speech (TTS) — turning written text into a spoken voice — by roughly 70 percent.

What's actually new in the lineup

Qwen-Audio-3.1 refreshes three existing models — ASR, TTS and Realtime — and adds two new ones built for more specialised jobs:

• ASR-Next handles "audio understanding": it can separate multiple speakers in a recording, label who said what, add timestamps, caption background sounds, and even flag the emotional tone of a clip.
• TTS-Next is built for "audio creation," pairing a language model with a diffusion model so it can generate a voice, sound effects and background ambience together in a single pass — aimed at audiobooks, podcasts, games and video production.
• The upgraded Realtime model (marketed as Realtime Plus) now supports a 262,000-token context window, full-duplex conversation (both sides can talk and be heard at once, like a real phone call instead of a walkie-talkie), function calling, and voice cloning.
• The base ASR model claims support across 30 languages and more than a dozen Chinese dialects, with a first-word response time of around 160 milliseconds — fast enough that a live transcript barely lags behind the speaker.

Alibaba Group Chairman Joe Tsai framed the broader push at the conference in typically expansive terms:

"The theme of this year's Apsara Conference is 'Intelligence Goes Beyond.' This is the true meaning of 'Intelligence Goes Beyond': guiding AI from technological breakthroughs toward value creation."

Cheaper voice AI is about as concrete a version of "value creation" as a cloud vendor can offer — most developers won't remember the slogan, but they'll notice the invoice.

Why this matters more in India than it sounds

Voice interfaces solve a real problem for India specifically: a huge share of internet users are more comfortable speaking than typing, especially outside English-first metros. That's why customer support bots, banking IVR systems, and edtech apps here have leaned so heavily on speech AI — and why cost has been the thing quietly killing a lot of those projects before they reach real scale. Running ASR and TTS across tens of millions of calls in a country where margins are thin makes a 95 percent price drop on speech recognition genuinely decision-changing for an Indian startup's unit economics, not just a nice-to-have.

Alibaba's wider Qwen ASR lineup has previously listed Hindi among its supported languages, though full coverage of India's other major languages — Tamil, Bengali, Telugu, Marathi and the rest — isn't something Alibaba has clearly committed to for this release. That gap is exactly where homegrown efforts like the government's Bhashini language mission and India-based voice AI startups still have room to compete. Code24 covered a similar move recently when Meta launched its Muse Voice Transcribe tool with support for five Indian languages — between that and Qwen's price cuts, Indian developers suddenly have more affordable voice AI options than they did a month ago.

There's a less comfortable side to this too. Indian banks, fintechs and health-tech firms are already navigating the Digital Personal Data Protection (DPDP) Act and RBI's data-localisation rules, which restrict where sensitive customer data can be processed and stored. Routing customer voice data — often the most personal kind of data a company collects — through a Chinese cloud provider is a different compliance conversation than using a US or India-based one, regardless of how good or cheap the model is. Any Indian company tempted by Qwen's pricing should be asking where the audio actually gets processed before they wire it into a production call centre.

The takeaway

Price wars in AI infrastructure usually get framed as a US-China story, but the actual winners are often builders in markets like India, where margins are tight and multilingual support was never optional. Cheap, capable voice AI won't fix India's compliance questions on its own — but it does mean more Indian teams can now afford to try building the voice product they'd shelved as too expensive six months ago.