Phonon vs Whisper Large v3 Turbo: Two Halves of an Offline Voice Loop

Comparing Phonon 31M with Whisper Large v3 Turbo is a category error, and it is the most useful category error in speech engineering right now. Phonon 31M, released by Gradium on 20 October 2026, turns text into audio. Whisper Large v3 Turbo, released by OpenAI on 1 October 2024, turns audio into text. They do not compete. They are the two ends of the same loop, and both of them are small enough and fast enough to run on hardware you own.

There is a second reason the pairing confuses people. Both publish a number called WER, and the two numbers mean opposite things. For Whisper, word error rate measures how often the transcription is wrong. For Phonon, it measures how often a transcriber mishears the speech the model generated. A lower number is better in both cases, but they are not the same measurement, and putting them in one column is how bad comparisons get made.

One listens, one speaks

Whisper Large v3 Turbo is a speech recognition model. It keeps the full encoder of Whisper large-v3 and shrinks the decoder from 32 layers to four, producing 809 million parameters instead of 1.55 billion. It covers 99 languages, auto-detects language, and is not trained for translation. OpenAI's documentation describes it as roughly eight times faster than the large model with a small accuracy cost; the company did not publish a precise multiplier on the model card, so any exact speed figure is third-party testing.

Phonon 31M is a speech synthesis model. It takes text and a voice and produces 24 kHz mono audio through a flow-matching transformer over Mimi latents. It is 31 million parameters excluding the tokenizer, ships as a 64.5 MB q8 GGUF, and is Apache 2.0. It has no listening capability at all — it cannot transcribe, diarise or detect language.

The only thing they share is a design goal: both are built to run without a data centre. Whisper Turbo is the fast, cheap member of a family that has become the default open ASR baseline. Phonon is an attempt to make the synthesis half equally cheap.

The same three letters, two different measurements

Whisper Large v3 Turbo reports a mean WER of 7.83 on the English Open ASR Leaderboard, against 7.44 for the full large-v3. On LibriSpeech test-clean the figures are around 2.0%. Those numbers describe how often the model writes the wrong word when it hears clean speech.

Phonon 31M reports 0.51% on Seed-TTS and 1.5% on a harder set in its model card, or 1.70% on the same hard set in the release blog. Those numbers describe how often whisper-large-v3 writes the wrong word when it hears Phonon's output. Notice that the scorer is Whisper itself. Gradium's methodology is circular in an unavoidable way: to measure a TTS model you need an ASR model, and the ASR model's own error rate sets a floor on how good the TTS can look.

That is why a TTS WER of 0.5% and an ASR WER of 7.8% are not comparable, and why a TTS WER below about 1% should be read as "the transcriber had no trouble," not "the speech is nearly perfect." Human listeners and automatic scorers disagree, especially on prosody, which no WER captures.

• Task — text to speech for Phonon 31M; speech to text for Whisper Large v3 Turbo.

• Parameters — 31M excluding the tokenizer for Phonon 31M; 809M for Whisper Large v3 Turbo.

• Languages — English only for Phonon 31M; 99 for Whisper Large v3 Turbo.

• Output — 24 kHz mono audio for Phonon 31M; text with timestamps for Whisper Large v3 Turbo.

• Reported accuracy — 0.51% Seed-TTS WER for Phonon 31M, measured by Whisper; 7.83 mean English WER for Whisper Large v3 Turbo on the Open ASR Leaderboard.

• Licence — Apache 2.0 for Phonon 31M; MIT for Whisper Large v3 Turbo.

Sizing an offline voice loop

If the goal is a voice assistant that works with the network off, the bill of materials is three models, not two. Speech recognition, language generation, and speech synthesis.

Whisper Large v3 Turbo is 809 million parameters, about 1.6 GB of weights at 16-bit precision. GPU inference typically reserves several gigabytes of working memory for activations and batching; quantized GGUF builds shrink the footprint further at a small accuracy cost. It is not a phone-class model in its Turbo form, though smaller Whisper variants and distilled versions are.

Phonon 31M is 64.5 MB. That is roughly one twenty-fifth of Whisper Turbo's weight file, and it runs on a CPU core, an iPhone Neural Engine or a Raspberry Pi.

The language model in the middle is the variable. A 1B to 4B instruct model quantized to four bits fits in a couple of gigabytes and is the usual choice for an offline assistant. The full loop — Whisper Turbo, a small instruct model, Phonon — lands in a range that a laptop or a high-end phone can hold, with the ASR as the heaviest component.

That asymmetry is the practical insight. The half of the loop people worry about — synthesis, the part that has to sound human — is now the cheap half. Recognition is the expensive one.

The latency budget

Phonon's vendor figures put first audio at 5 to 7 ms on a MacBook Air M5 core and 9 ms on an iPhone 16 Pro, with throughput between 11.6 and 41 times real time depending on which of Gradium's documents you read. Whisper Turbo's speed depends heavily on hardware and backend: OpenAI describes it as roughly eight times faster than large-v3, and community benchmarks with quantized builds report tens of times real time on Apple silicon and much slower figures on CPU.

For a conversational loop, the recognition step dominates. A three-second utterance takes a few hundred milliseconds to transcribe on a laptop GPU and close to real time on a weak CPU. The synthesis step is effectively instantaneous by comparison. Teams that optimise the TTS first are optimising the wrong end.

Language coverage is the hard constraint

Whisper handles 99 languages. Phonon 31M handles one, and Gradium's card warns that other languages are mispronounced. A multilingual voice assistant therefore cannot be built from this pair: recognition works in Japanese, synthesis does not.

Gradium's answer is the Phonon 40M family, which covers English, French, German, Spanish and Portuguese. That is five languages, not 99, and the 40M checkpoints are CC BY-NC 4.0 — not usable in a commercial product without a separate agreement. The multilingual gap between open ASR and open TTS is one of the widest in the stack, and nothing released this month closes it.

Where they meet in a real product

The natural application is a transcription-and-playback tool: record a meeting or a voice note, transcribe it with Whisper, then read a summary aloud with Phonon. Both run locally, both are permissively licensed, and neither sends audio to a third party. A medical dictation tool, a field recorder, an accessibility app for deaf and hard-of-hearing users, or an offline language-learning toy all fit the same shape.

The second application is a voice agent with a local first pass and a cloud fallback. Whisper transcribes on the device; if confidence is low, the audio goes to a hosted ASR. The language model drafts a reply; if it needs a bigger model, the request routes out. Phonon speaks the result; if the device is thermally throttled, a hosted TTS takes over. Each stage has a local and a remote option, and the product decides which to use per request.

The text layer between them

Whisper produces a transcript, Phonon consumes a script, and something has to turn one into the other. That something is a language model, and it is the only part of the loop that benefits from being large.

OrcaRouter does not host either of these models — Whisper Large v3 Turbo runs through your own inference stack, and Phonon runs on the device — so neither appears in our catalogue. What the platform covers is the middle: one API for more than 200 language models at provider list price with 0% markup, automatic failover across providers, and a routing DSL that can send a cheap model the routine turns and a stronger one the hard ones. In an offline-first design, that layer is the part that stays online.

What to actually compare

If you are choosing a speech model, compare Phonon against other TTS systems and Whisper Turbo against other ASR systems. If you are building a product, compare the loops: total memory, total latency, total licences, total languages. On that measure the pair is coherent for English-only, offline, permissive-licence work and incoherent for anything multilingual or commercial in the 40M tier.

The mistake to avoid is reading Phonon's 0.51% WER next to Whisper's 7.83% and concluding that synthesis has surpassed recognition. The numbers are produced by different processes measuring different things, and one of them is measured by the other. The useful question is not which model is better. It is whether the loop you can assemble from open parts this month is good enough for the product you have in mind — and for English, offline, on a laptop, the answer just changed.