Alibaba Releases Qwen-Audio 3.1 and Cuts Voice API Prices by Up to 95%

Building an AI voice app may now cost less, whether it transcribes speech or talks back. Alibaba has released Qwen-Audio 3.1, a five-model lineup for speech recognition, speech generation and live conversation. Its reported API price cuts reach up to 95%, though the company’s cloud catalog does not yet show a 3.1 text-to-speech model in its recommended listing.

The price changes vary sharply by task. Text-to-speech costs fall about 70%, live conversation about 85%, and speech recognition up to 95%, according to The Decoder. Those percentages describe API price reductions, not the cost of running a finished voice service. A developer weighing a switch would still have to consider how much transcription, generated speech and live interaction an app actually uses.

From a recording to a response

The lineup divides voice work into three familiar jobs: recognizing what someone said, generating spoken audio and holding a live exchange. The standard speech-recognition model is designed to handle multiple languages and dialects. It also automatically removes filler words and repetitions. That may produce a cleaner transcript for reading, but an app that needs a verbatim record would have to check what the cleanup changes.

ASR-Next goes beyond the words. It identifies different speakers, assigns timestamps and detects emotions and sounds such as ambient noise or machinery. Those features target recordings where the question is not just what was said, but who said it and what else was audible. They also raise a different test for developers: whether the extra labels are useful and accurate on their own recordings.

More control over the sound

On the output side, Qwen’s text-to-speech model can generate speech in multiple languages and transfer a voice across languages. A text instruction can steer its emotion, speed and delivery style. That gives a creator more to specify than the words alone: the same script can call for a different pace or tone without changing the text being read.

TTS-Next has a broader job. It pairs a language model with a diffusion-based generation method to produce voice, sound effects and background audio in one pass. For a narrated piece, that would bring elements normally handled as separate audio outputs into one generation step. The feature description alone, however, does not show how much editing a finished result would need.

The live-conversation model tackles a different problem: waiting for a turn to end. It can listen while speaking and accept an interruption, rather than requiring a rigid back-and-forth. Qwen says it slows down and responds more empathetically when it detects a low mood, but that describes intended behavior—not evidence it will read someone’s feelings correctly in every conversation.

What the cloud catalog shows

Alibaba Cloud’s recommended-model catalog lists two Qwen-Audio 3.1 speech-recognition entries: one for streaming audio and one for file transcription. It also lists Qwen-Audio 3.1 Realtime Plus for speech-to-speech conversation. Developers can use those model names to identify endpoints for testing transcription or a live voice app, rather than treating the five-model lineup as a single endpoint.

Its recommended text-to-speech listing, by contrast, names Qwen-Audio 3.0 TTS Plus, not a 3.1 model. The catalog is not a complete inventory or release schedule, so the contrast does not settle when every announced capability will be accessible. Still, it makes the rollout worth checking by endpoint: the published lineup and catalog entries are not identical.

The release therefore offers two things to judge separately: lower reported API rates and a wider set of voice capabilities. The cuts could be substantial for an app whose spending is concentrated in transcription or live calls. For anyone choosing a provider, though, the decision still turns on the available endpoint and how well it handles the voices, interruptions or recordings their app depends on.