Qwen Audio 3.1: Five New Speech Models and Price Cuts up to 95%
What happened
On September 23, 2026, Alibaba's Qwen team released Qwen-Audio-3.1: a lineup of five models covering speech recognition (ASR), text-to-speech (TTS) and real-time voice interaction. Alongside the release, Alibaba cut audio pricing sharply — TTS down about 70%, Realtime roughly 85%, and ASR by up to 95%, as reported by The Decoder.
The price move is the part worth paying attention to. Speech has been the one modality where per-minute costs still quietly shaped what teams were willing to build — transcribing an entire support-call archive or giving every user a real-time voice agent was a budget decision, not a technical one. A 95% cut on ASR changes which of those ideas survive a spreadsheet.
The five models, and what each is actually for
ASR — the baseline speech-to-text model. Improved multilingual and dialect recognition, and it automatically cleans up filler words and repetitions. That last detail matters more than it sounds: raw transcripts full of "um", "you know" and restarts are the reason so many transcription pipelines need a second cleanup pass. Removing it at the source removes a step.
ASR-Next — the one to look at for meetings and interviews. It adds multi-speaker identification with timestamps, and detects emotion, ambient sounds and machine noise. Speaker-attributed, timestamped output is the difference between a wall of text and something you can actually navigate. Ambient and machine-noise detection is a nod at industrial and monitoring use cases, not just conversation.
TTS — multilingual synthesis with natural cross-language voice transfer, meaning one voice can carry across languages. Delivery is controlled through plain text prompts rather than tuning parameters: Qwen's own example is "Read this with a sharp, commanding tone, demanding respect." If you've fought with SSML tags to get a line read correctly, prompt-controlled emotion, speed and style is a real ergonomic shift.
TTS-Next — pairs a language model with a diffusion approach to generate voice, sound effects and background audio in a single pass. This is the one aimed at production rather than narration: a scene rather than a read-aloud. Single-pass generation means you're not stitching a voice track to a separately-sourced ambience bed and hoping they sit together.
Realtime — supports simultaneous speaking and listening with instant interruption. Qwen also says that when it detects a low mood, it responds more slowly and with more empathy. Barge-in handling is the make-or-break feature for voice agents; a model that can't be interrupted mid-sentence feels broken to callers regardless of how good its answers are.
Full details are on the Qwen blog and Qwen Cloud.
What the price cuts actually mean
Percentages off a list price are easy to over-read, so it's worth being precise about which cut matters for which workload:
| Cut | Model | Workload it unlocks |
|---|---|---|
| up to 95% | ASR | Bulk, backward-looking transcription — call archives, podcast back-catalogues, every meeting you've ever recorded |
| ~85% | Realtime | Always-on voice agents, where cost scales with conversation minutes and never stops |
| ~70% | TTS | Per-item narration — article read-alouds, notifications, generated video voiceover |
The ASR cut is the one that changes project viability rather than margin. Transcribing a large historical archive is a fixed, one-time bill: at old prices it needed a business case, and at a twentieth of the price it's often just an afternoon. Realtime's cut is different in kind — it's a recurring cost per minute of conversation, so an 85% reduction changes what per-user economics look like for a voice product you intend to run indefinitely.
One caution: the honest way to read any price cut is against the incumbent you'd otherwise pay, not against the previous Qwen price. If you're already on a competing ASR provider, the number to run is the new Qwen per-hour rate against your current per-hour rate on your audio — accented speech, overlapping speakers, domain jargon. Accuracy differences of a few percent on your actual material will dominate a list-price comparison, because the cost of fixing bad transcripts is human time.
Can you try it on AIWITH.CHAT?
Not the audio models — no. AIWITH.CHAT is a text-and-image chat product, and there are no ASR, TTS or real-time voice models on it today, Qwen's or anyone else's. If you specifically want Qwen-Audio-3.1, the place to get it is Qwen Cloud or Alibaba Cloud's API directly. Saying otherwise would waste your time.
What is worth knowing is that audio work is rarely the whole job. Transcription is a means to something else — a summary, a set of action items, a script, a translated version, a social post cut from a recording. That second half is a text-model job, and it's the half where switching models actually changes the output quality:
| Model | Best for the post-transcript half |
|---|---|
| GPT-6 Astra | Long transcripts where you need structure pulled out reliably |
| Claude Opus 5.5 | Turning a raw transcript into something publishable |
| Gemini 3.1 Pro | Very long recordings held in one context |
| DeepSeek V4.1 Flash | High-volume, cheap passes over many transcripts |
| Qwen Image / Nano Banana Pro | Thumbnails and visuals for the finished piece |
All of those sit in one dropdown under a single $9.9/month plan, with no separate per-model subscriptions — so running the same transcript through Opus 5.5 and GPT-6 Astra to see which writes the better summary costs you nothing extra.
See what's included on AIWITH.CHAT →
Is it worth switching?
If you run bulk transcription: yes, price it out this week. A 95% ASR cut is large enough that it beats almost any optimization you'd otherwise attempt, and archive transcription is a one-time bill that either clears your threshold or doesn't. Benchmark on your own worst audio, not on a clean sample.
If you're building a voice agent: the Realtime model's simultaneous listen-and-speak with instant interruption plus an 85% cut is a genuinely strong combination, and barge-in is the feature most voice stacks get wrong. Worth a prototype.
If you just need occasional narration or the odd transcript: probably not urgent. At low volume you're optimizing a bill that's already small, and whatever you're using works. Revisit when volume grows.
If audio is only the front half of your workflow: the model you pick for the second half matters at least as much. Cheap, accurate transcripts of a meeting are still just text until something turns them into a decision.
One-sentence summary
Qwen-Audio-3.1 brings five speech models — including speaker-attributed ASR with emotion detection and a single-pass voice-plus-sound-effects TTS — and cuts audio prices up to 95%, which mostly changes the maths on bulk transcription and always-on voice agents; the audio models live on Qwen Cloud, not on AIWITH.CHAT, but the text models that turn transcripts into finished work do.