Phonon vs Qwen Audio 3.0 TTS: 7 ms on Your Laptop or 200 ms in the Cloud

A support agent finishes a sentence and the voice assistant has to answer. On Phonon 31M, Gradium's 64 MB on-device model released on 20 October 2026, the first audio sample can arrive about 7 milliseconds after the text is ready, because nothing leaves the machine. On Qwen Audio 3.0 TTS, Alibaba's hosted system from July 2026, the first packet is documented at under 200 milliseconds — and that figure starts after your request has crossed the network to a data centre in Beijing or Singapore.

That gap is the whole comparison in miniature, but it is not the whole decision. Qwen Audio 3.0 TTS speaks 16 languages and 20 Chinese dialects at 48 kHz, takes emotion and non-verbal tags inline, and clones a voice from a reference clip. Phonon 31M speaks English at 24 kHz in 37 fixed voices and runs on a Raspberry Pi. Choosing between them is choosing what kind of system you are building, not which model sounds nicer in a demo.

What each one actually is

Phonon 31M is one checkpoint in Gradium's open-weight family. It is 31 million parameters excluding the audio tokenizer, ships as a 64.5 MB q8 GGUF, outputs 24 kHz mono, and is licensed Apache 2.0 — the only checkpoint in the family that is free for commercial use. The companion Phonon 40M models cover five languages but sit under CC BY-NC 4.0, so they are not an option for a shipped product without a separate agreement. The runtime is Rust, with no PyTorch at inference, and builds exist for CPU, Core ML, Snapdragon NPUs and WebAssembly.

Qwen Audio 3.0 TTS is not a downloadable model. Alibaba's Qwen team announced it on 20 July 2026 at WAIC in two API tiers: Flash, tuned for real-time interaction, and Plus, tuned for quality. Both are reached through Alibaba Cloud Model Studio. The documentation for the Flash endpoint lists a first-packet latency of under 200 ms, a rate limit of 180 requests per minute, and a price of 1 CNY per 10,000 characters in Beijing, or 1.12413 CNY in Singapore. There are no weights, no local runtime, and no offline mode.

The latency difference is not 193 milliseconds

Subtracting 7 from 200 gives a number that understates the gap and overstates its importance, depending on what you are building. For a turn-taking voice interface, the end-to-end budget is roughly 800 ms before conversation starts to feel wrong. A cloud TTS call spends its 200 ms after a network round trip that can be 20 ms on a good wired connection and 150 ms on mobile. Add TLS setup on a cold connection, audio buffering, and the tail latency that a 180 RPM rate limit implies under load, and the p95 experience is materially worse than the headline.

The local path has a different shape. Phonon's own README lists 5 ms time-to-first-byte on one MacBook Air M5 core and 9 ms on an iPhone 16 Pro Neural Engine. There is no network tail, no rate limit, and no provider having a bad afternoon. What you give up is elasticity: a phone that is thermal-throttling runs Phonon slower, and a browser tab competing with a video call gets less CPU. The vendor's Raspberry Pi 5 figures — 68 ms and 2.3 times real time in the README, 38 ms and 3.7 times in the release blog — show how far that floor can drop on weak hardware.

This is also where the numbers deserve a caveat. Gradium's release blog and its runtime README disagree on throughput by more than a factor of two for the same device, and neither has been independently reproduced. Treat 7 ms as an order of magnitude, not a specification.

Languages and voices: breadth versus a fixed cast

Qwen Audio 3.0 TTS covers 16 languages, including Japanese, Korean, Arabic, Vietnamese, Malay and Filipino, plus 20 Chinese dialects. It supports voice cloning from reference audio and can output a different language from the one in the reference clip, which is the feature that makes localisation pipelines interesting. Output is 48 kHz, and the platform allows up to three minutes of audio per synthesis.

Phonon 31M is English-only, and Gradium's model card is blunt about it: other languages are mispronounced. It carries 37 built-in voices, with more in a separate voice catalogue and support for custom voices in .safetensors format. The multilingual Phonon 40M checkpoint reaches five languages, but its non-commercial licence removes it from most product decisions.

• Languages — English only for Phonon 31M; 16 languages plus 20 Chinese dialects for Qwen Audio 3.0 TTS.

• Sample rate — 24 kHz mono for Phonon 31M; 48 kHz for Qwen Audio 3.0 TTS.

• Voice sources — 37 built-in plus a custom-voice catalogue for Phonon 31M; reference-audio cloning for Qwen Audio 3.0 TTS.

• Maximum utterance — not documented for Phonon 31M; up to three minutes per synthesis for Qwen Audio 3.0 TTS.

• Where it runs — CPU, NPU, browser or phone for Phonon 31M; Alibaba Cloud Model Studio only for Qwen Audio 3.0 TTS.

• Licence — Apache 2.0 for Phonon 31M; commercial API terms for Qwen Audio 3.0 TTS.

Control: tags and cloning versus a fixed cast

Qwen's expressiveness story is built on inline control. The model accepts fine-grained non-verbal tags — breaths, laughs, sighs — and natural-language style instructions, so a script can specify not just what is said but how. Launch coverage highlighted the shift from "can speak" to "can perform," and the Plus tier was ranked first on the Artificial Analysis Speech Arena with an Elo of 1,237, ahead of Gemini 3.1 Flash TTS and ElevenLabs v3. That is an independent leaderboard rather than a vendor table, though the entry was submitted by Alibaba and Elo on a curated arena rewards style as much as intelligibility.

Phonon's control surface is deliberately narrow. You pick a voice, you send text, you get audio. There are no emotion tags and no style prompt. For an interface that reads short, predictable utterances — notifications, navigation, form confirmation — that is enough, and the absence of a style layer is one reason the model is 31 million parameters instead of billions. For narration, advertising or character work, it is disqualifying.

Neither system solves the consent problem. Qwen's cloning accepts a reference clip; Phonon accepts custom voice files. Both leave it to the integrator to establish that the speaker agreed.

What it costs at volume

Phonon 31M has no per-character price. You pay for hardware you already own and for the electricity to run a model that generates 26 to 41 times faster than real time on a laptop core. At small scale that is effectively free; at large scale it becomes a mobile battery and thermal budget, which is a real cost but a different one.

Qwen Audio 3.0 TTS Flash bills 1 CNY per 10,000 characters in Beijing. Ten million characters a month is 1,000 CNY, roughly US$140 at a round seven yuan to the dollar — less than most teams spend on logging. The number that matters is not the monthly bill but the shape of it: cost scales with usage, latency is rented, and the model can be repriced or deprecated by someone else. That is the trade for 48 kHz and sixteen languages.

OrcaRouter does not host either of these models, and nothing here should suggest otherwise: Phonon runs on your device, and Qwen Audio 3.0 TTS is reached through Alibaba's own API. Where routing does earn its place is one layer up. A voice assistant's reply text usually comes from a language model, and that model is a commodity you can move. One key covers more than 200 models at provider list price with 0% markup, so a price cut or a new cheap model is usable the same day, and automatic failover keeps the text half of the loop alive when a single provider degrades. The speech layer can stay exactly where you put it.

Where each one wins

Choose Phonon 31M when the audio must be generated on the device, when the language is English, when a fixed voice set is acceptable, and when a 64 MB download and an Apache 2.0 licence are worth more than stylistic range. Offline assistants, accessibility tools, games, kiosks and browser features are its natural habitat.

Choose Qwen Audio 3.0 TTS when you need languages other than English, 48 kHz output, emotional performance, voice cloning, or utterances longer than a sentence or two — and when a network dependency is acceptable. Customer-facing voice agents in multiple markets, dubbing, and narrated content are its natural habitat.

The hybrid most teams will end up building

The two systems are not mutually exclusive, and the interesting architecture uses both. Run Phonon 31M as the local default for short, predictable, English responses — the ones where latency is felt. Fall back to Qwen Audio 3.0 TTS when the response is long, when the user has switched language, or when the device cannot sustain the local model. The switch is a policy decision, not a rewrite, because both take text and return audio.

What makes that architecture work is that the fallback is not a second-best experience. Qwen's 48 kHz output and style tags are better than Phonon's on almost every axis except latency and privacy. The reason to default local is not quality; it is that 7 milliseconds beats a network round trip when the user is waiting.

Gradium's release is three days old and its benchmarks are still vendor-run. Alibaba's model is three months old, has an independent arena placement, and is priced per character. The decision between them will not be settled by a WER table. It will be settled by whether your product can tolerate a data centre in the loop.