Qwen-Audio-3.1: A Full Audio Stack for Automation Builders
September 25, 2026 · AI Automators
Qwen has expanded its audio family with Qwen-Audio-3.1, a set of models spanning speech recognition, text-to-speech, realtime voice, and audio understanding. For anyone building automations that touch audio — meeting transcripts, call analysis, voiceovers, or voice agents — this is worth a look, mostly because of what it does and partly because of how the pricing has moved.
The announcement describes five models across four jobs: understanding, generation, interaction, and creation. Two of them are new: TTS-Next for audio creation and ASR-Next for audio understanding. Alongside the release come price cuts described as roughly 70% off TTS, around 85% off Realtime, and up to 95% off ASR. Those numbers come from Qwen's own messaging, so treat them as list-price claims rather than a guarantee about your specific usage, but the direction is clear: audio processing is getting cheaper to run at volume.
What each model actually does
The clearest documented piece is the ASR model on QwenCloud, listed as `qwen-audio-3.1-asr-flash-filetrans`. It's an end-to-end offline speech recognition model built for meeting transcription, content production, and call analysis. It supports multilingual input and multiple Chinese dialects, and it ships with the features that make transcripts usable in practice: speaker separation, punctuation prediction, text normalization, and hot-word or context enhancement so you can bias it toward names and jargon it might otherwise miss. Qwen says it stays stable in noisy conditions, which matters more than headline accuracy for real-world call recordings.
One detail from the announcement stands out for automation: the ASR does "native polishing" that auto-removes fillers and repetitions to produce cleaner, more logical transcripts. If you've ever piped raw ASR output into a summarization step, you know how much of the cleanup work you normally have to do yourself.
ASR-Next goes further into understanding rather than plain transcription. It supports multi-speaker recognition with speaker labels, timestamps, and aligned transcripts, and it claims to interpret emotion, ambient sounds, and machine noises — enabling sound captioning, event localization, and audio question-answering. That's a different category of tool: less "turn speech into text" and more "answer questions about what happened in this recording."
On the generation side, TTS handles multilingual and dialect synthesis with cross-lingual voice transfer, and lets you control emotion, speed, and style through instructions. TTS-Next combines a language model with a diffusion framework to produce voice, sound effects, and background audio in a single pass — aimed at audiobooks, podcasts, games, and ads. Realtime supports simultaneous speak-and-listen with interruption at any point, and the announcement notes it can slow down and respond more empathetically when it senses a low mood. How well that empathy claim holds up in practice is exactly the kind of thing you'd want to test before relying on it.
Fitting it into an automation stack
The ASR endpoint is built for asynchronous, file-based transcription. The documented API takes a list of `file_urls`, runs with the `X-DashScope-Async: enable` header, and lets you specify a channel. That async, URL-based design fits batch workflows well: drop recordings in storage, hand the URLs to the API, and pick up results when they're done. It's a natural fit for an n8n or Make pipeline where a new file in a bucket triggers a transcription job, then feeds the polished transcript into a summarizer or a database. Zapier can orchestrate the same pattern if you prefer a hosted builder.
The QwenCloud listing also shows the platform features around the model: structured outputs so you can force JSON, batch processing to cut cost, function calling, context caching, and fine-tuning. Structured outputs and batching are the two most relevant for automation — the first keeps downstream parsing predictable, the second matters when you're transcribing large archives. Published ASR pricing on the page is $0.15 per 1M input tokens and $0.47 per 1M output tokens, with a rate limit of 600 requests per minute.
Where does this sit against alternatives? For pure transcription, it competes with tools like Deepgram, AssemblyAI, and OpenAI's Whisper-based endpoints. Qwen's pitch leans on multilingual and dialect coverage — particularly regional Chinese — plus the built-in transcript polishing and speaker handling. If your workloads are English-only, the established Western providers may still be the safer default; if you're processing Chinese-language or multilingual audio, the dialect support is a genuine differentiator worth benchmarking against your own recordings.
A few practical cautions. The full lineup is rolling out, and the announcement notes "more APIs coming soon," so not every model may be callable yet through the same interface. Pricing and rate limits shown are the published values and can change. And the more advanced claims — emotion detection, sound event reasoning, empathetic realtime responses — are the ones most worth verifying against real inputs before you build them into anything customer-facing.
For teams already leaning on OpenAI or Claude for the text side of their pipelines, Qwen-Audio-3.1 is a plausible way to handle the audio front-end more cheaply, especially for multilingual work. If you want help wiring transcription, TTS, or voice agents into a working automation, browse the provider directory to find people who can put it to work.