Alibaba's Qwen Audio 3.1 Cuts ASR Prices 95% and Ships Five Models — The Voice AI Stack Just Got Cheaper

The economics of building voice-first AI features just shifted. Alibaba’s Qwen Audio 3.1 release cuts ASR prices by up to 95%, makes real-time speech processing 85% cheaper, and introduces a new TTS architecture that generates voice, sound effects, and background audio in one model pass. Together, the five models in this release cover the full audio AI stack at price points that change what’s viable to build.

The Five Models

ASR (Automatic Speech Recognition): The updated base model improves multilingual and dialect recognition and adds automatic filler-word removal — “um,” “uh,” and repetitions are stripped from transcriptions without a post-processing step. Price cuts of up to 95% bring this to the lowest cost-per-minute in Alibaba’s audio lineup.

ASR-Next: Adds multi-speaker identification with timestamps — the model labels who said what and when — alongside detection of emotions, ambient sounds, and machine noise. For meeting transcription, call center analysis, or any scenario where speaker attribution and context matter, ASR-Next goes further than the base model.

TTS (Text-to-Speech): Handles multilingual synthesis with natural cross-language voice transfer. The distinguishing feature: users can “control emotion, speed, and style through simple text prompts” — no voice cloning required to adjust how the output sounds.

TTS-Next: The architectural shift in this release. Instead of a standalone TTS pass followed by separate sound design, TTS-Next pairs a language model with a diffusion process to generate voice, sound effects, and background audio in a single inference pass. For applications that need narration, ambient sound, and effects together — interactive audio experiences, game dialogue, or audiobook production — this collapses multiple API calls into one.

Real-time Model: Supports simultaneous speaking and listening with instant interruption handling. The model adjusts emotional tone based on conversational mood, making it suitable for voice agents that need to respond naturally to user state rather than treating every turn as a neutral information exchange.

The Price Cuts in Context

At 95% ASR reduction, the unit economics argument against voice-first features in AI applications largely disappears. The previous calculus — “voice is too expensive for anything but high-value interactions” — no longer holds for teams using Qwen Audio APIs.

This is competitive with, and in some tiers undercuts, OpenAI Whisper API pricing and ElevenLabs TTS. The practical effect is downward price pressure across the voice AI market.

What the TTS-Next Architecture Means

The conventional pipeline for generating AI audio content runs: text → TTS model → separate sound design → audio mixing. TTS-Next collapses the first two steps. The diffusion component generates ambient context alongside voice — if the text says “whispered instructions in a noisy market,” the model generates both the whispered voice and appropriate market sound texture from the same inference call.

This matters less for simple narration use cases (where you just need a voice) and more for interactive audio applications, game dialogue systems, and anywhere immersive context is part of the design. It also opens the door to novel applications that previously required audio production skills to execute — the model handles the sound design judgment that would otherwise need a human.

Content created with AI assistance and reviewed for accuracy.