Voice AI just got dramatically cheaper. On September 24, 2026, Alibaba's Qwen team announced Qwen-Audio 3.1 — a five-model voice stack that drops speech recognition API pricing by up to 95%, text-to-speech by roughly 70%, and the realtime voice model by around 85%. If you've been holding off on voice-enabling your application because the costs were prohibitive, the math just changed.
Five Models, One Complete Stack
The release upgrades three existing models — ASR (speech recognition), TTS (text-to-speech), and Realtime — while introducing two brand-new additions. ASR-Next is built for audio understanding: it performs multi-speaker transcription with per-speaker labels and timestamps, detects emotions, identifies ambient and mechanical sounds, and supports full audio Q&A and reasoning. It's not just transcription — it's a full audio comprehension system.
TTS-Next goes in the opposite direction — audio creation. It pairs a language model with a diffusion-based approach to generate voice, sound effects, and background audio simultaneously in a single inference pass. Generating a podcast segment with ambient coffee-shop noise and natural speech no longer requires chaining three separate APIs.
The Pricing Story
The cuts are steep enough to reshape competitive dynamics. ASR pricing falls by up to 95%, which puts Qwen's offering well below what most enterprise vendors have been charging. The Realtime model — the one used for low-latency conversational applications — drops by approximately 85%. For context, Alibaba's own earlier Qwen3.8-Omni-Flash had already slashed audio input pricing by 98% in September, moving from $0.28 per hour of audio to under $0.01. Qwen-Audio 3.1 extends those reductions across the full product lineup.
This is a familiar pattern from Alibaba's cloud strategy: use aggressive pricing to gain developer adoption, then compete on ecosystem depth. The difference this time is that the underlying models are genuinely compelling. ASR-Next's emotion and ambient sound detection puts it in a different capability class from simple transcription engines.
What It Means for Builders
For developers building voice assistants, call-center automation, accessibility tools, or any audio-heavy application, the barriers just dropped significantly. Multi-speaker diarization with emotion detection — a feature that previously required either expensive specialized vendors or significant custom engineering — is now bundled into a standard API call.
TTS-Next's single-pass audio generation is particularly interesting for content creators: generating narration with appropriate background ambiance in one step, rather than compositing separate audio tracks in post, could meaningfully speed up production pipelines.
Alibaba has published pricing details and model documentation through the Qwen team's official announcement. The models are accessible through Alibaba Cloud's API platform and, for those who want to run locally, select open-weight variants are available on Hugging Face.
The broader takeaway is that audio AI is following the same deflationary trajectory that text AI did 18 months ago: prices collapsing, capabilities expanding, and the competitive moat for any vendor resting on model quality and developer experience rather than access controls.