Phonon vs Sesame Preview: 32× Fewer Parameters, the Same Codec, a Different Bet

Put Phonon 31M and Sesame Preview side by side and the first thing that matches is the plumbing. Both produce 24 kHz mono speech, both encode it through Kyutai's Mimi codec, and both are Apache 2.0. The second thing that matches is stranger: the open checkpoint behind Sesame Preview, Sesame CSM-1B, carries roughly a billion parameters. Gradium's Phonon 31M, released on 20 October 2026, carries 31 million. That is a 32-fold gap in model size for two systems that output the same sample rate through the same decoder.

The gap is not a story about efficiency winning. It is a story about two different bets on what conversational speech needs. Sesame's model spends its parameters on context — it is designed to continue a dialogue, reading prior turns to keep a voice consistent and a tone appropriate. Phonon spends almost nothing on context and everything on being small enough to run inside a browser tab, a phone app or a Raspberry Pi. Nineteen months separate their releases, and in that time the small end of the market moved a long way.

First, what "Sesame Preview" actually refers to

This matters because most comparisons get it wrong. Sesame Preview is the consumer product — the Maya and Miles voices at sesame.com, a demonstration of what the company calls voice presence. The open weights are a separate thing: Sesame CSM-1B, published on 13 March 2025 under Apache 2.0, a base generation model that Sesame explicitly says has not been fine-tuned on any specific voice. The preview's voices come from a fine-tuned variant that was never released.

So a listener comparing the Preview demo to Phonon's open checkpoint is comparing a curated product to a raw model. That is not a fair fight, and it is the fight most people are actually having. The fair comparison is CSM-1B against Phonon 31M: both downloadable, both Apache 2.0, both English-first, both open to inspection.

There is a second wrinkle in the Sesame card. The model is branded 1B, but the Hugging Face metadata lists 2B parameters in F32 tensors. The card does not explain the discrepancy. Either way, it is an order of magnitude above Phonon.

Same codec, opposite bets

CSM generates RVQ audio codes from text and audio inputs using a Llama backbone and a smaller audio decoder that emits Mimi codes. The architecture is a language model that happens to speak: it treats audio tokens as a sequence to be continued, conditioned on the conversation so far. That is why Sesame's documentation insists the model sounds best when given context, and why the recommended usage passes prior speaker segments rather than a single line of text.

Phonon is a flow-matching transformer predicting latents in 80 millisecond frames, decoded by Mimi. It builds on Pocket TTS. There is no dialogue state and no context window in the product sense — you give it text and a voice, and it speaks. The runtime is Rust with no PyTorch at inference, with CPU, Core ML, QNN and WebAssembly paths.

• Parameters — 31M for Phonon 31M; branded 1B for Sesame CSM-1B, with 2B listed in Hugging Face metadata.

• Output — 24 kHz mono for both.

• Codec — Mimi for both.

• Context — none for Phonon 31M; prior conversational segments recommended for Sesame CSM-1B.

• Languages — English for both, with Sesame warning that non-English is weak.

• Licence — Apache 2.0 for both.

What Sesame's parameters buy

Conversational presence is not a marketing phrase; it corresponds to a specific technical choice. A model that sees the previous turns can carry prosody across a sentence boundary, avoid resetting to a neutral read at every utterance, and keep a character voice stable through a long exchange. CSM-1B was built to do that, and the demo that made Sesame famous in March 2025 was convincing precisely because the voices sounded like they were listening.

The cost is that you need the context to get the benefit. A single-line synthesis through CSM-1B, with no prior segments, is not the experience the demo sold. It is a competent but unremarkable TTS read. Teams that download the checkpoint expecting the Preview's Maya voice discover this quickly: the voice is not in the weights.

Phonon makes the opposite trade. It has no notion of a conversation, so it cannot drift or reset — every utterance is independent and deterministic. For a notification, a navigation prompt or a game line, that is exactly right. For a companion character that needs to sound like it remembers the last sentence, it is not enough, and no amount of speed fixes it.

What nineteen months bought

The interesting number is not 32×. It is that a 31M model released in October 2026 claims English intelligibility in the same neighbourhood as systems two orders of magnitude larger. Gradium's own evaluation of an earlier, ~100M internal Phonon reported 1.48% WER on the English Seed-TTS set in April 2026, improving to 1.00% by May. The released 31M checkpoint is claimed at 0.51% on Seed-TTS and 1.5% on a harder set in the model card, with 1.70% on the hard set in the release blog.

Those are vendor-run numbers, produced with whisper-large-v3 as the transcriber and WavLM for speaker similarity, and they have not been independently reproduced. Sesame, for its part, never published a WER table for CSM-1B. The comparison that matters — same test set, same transcriber, both models — does not exist in public. Anyone who tells you the 31M model is more intelligible than the 1B model is extrapolating from two different harnesses.

What can be said is that the parameter gap no longer maps cleanly onto quality. In 2025, a billion-parameter model was a reasonable minimum for natural English TTS. In late 2026, a 64 MB file with 31 million parameters is credibly in the same conversation. That is the actual news in Gradium's release, and it is bigger than any single Elo figure.

Deployment: GPU versus a browser tab

CSM-1B is a PyTorch model. It runs on a GPU comfortably, on a CPU slowly, and it wants memory in the gigabytes rather than megabytes. Hugging Face Transformers added native support in May 2025, with batched inference and CUDA graph compilation, which tells you the intended deployment: a server, serving many requests.

Phonon 31M is 64.5 MB quantized and runs through a Rust runtime with bindings for Python, Swift, JavaScript and C. It starts in single-digit milliseconds on an Apple laptop core, around 9 ms on an iPhone 16 Pro Neural Engine, and 68 ms on a Raspberry Pi 5 in the README's figures — 38 ms in the release blog. It can run in a browser tab through WebAssembly. It cannot serve a thousand concurrent users from one GPU, because that is not what it is for.

This is the cleanest way to choose between them. If the voice lives in a data centre and serves many users, CSM-1B's architecture is a better fit. If the voice lives in the user's device and serves one person, Phonon's is.

Voice identity and consent

Sesame's open model produces a variety of voices but is not a cloning system out of the box; the card frames it as a base generation model and points to context as the way to keep a voice consistent. Gradium ships 37 built-in English voices with Phonon 31M, a larger catalogue at gradium/phonon-voices, and support for custom voice files in .safetensors format.

Both vendors publish acceptable-use language prohibiting impersonation and deception. Neither publishes a technical provenance mechanism — no watermark, no signed metadata, no detection API — in the material available as of 23 October. For anyone deploying either model in a customer-facing product, that is an open compliance question, not a solved one.

Where the Preview still wins

If the product is a conversational companion — a character that talks with a user over many turns and needs to sound like the same person throughout — Sesame's approach remains the more direct fit. The context conditioning is the feature, and Phonon has no equivalent. The Preview also has a polish that comes from a curated fine-tune and a curated voice cast, which no open checkpoint reproduces.

If the product needs speech on a device, in a browser, offline, or at a latency a network cannot match, Phonon is the more useful release. It is also the more inspectable one: the runtime is on GitHub, the weights are a single 64 MB file, and the licence is Apache 2.0 with no gate.

The part neither model handles

Sesame Preview has no public API, which means a team that wants its voice in a product cannot simply call it. The demo is a demonstration. Building on Sesame's research means running CSM-1B yourself, and that brings back the GPU and the context assembly.

Phonon handles synthesis and nothing else. In a real assistant, the text it speaks is produced by a language model, and the conversation state that Sesame's model would consume has to be assembled somewhere. OrcaRouter is not a speech provider and does not host either of these models — Phonon runs locally and Sesame's weights run on your hardware. What it does cover is the text layer those systems depend on: one endpoint for more than 200 models, provider list price passed through with 0% markup, and failover across providers so a single outage does not silence the assistant. The voice can stay on the device; the thinking does not have to.

The open question

Sesame's bet was that conversational context is what separates synthetic speech from a voice you believe. Gradium's bet is that most speech in software does not need a conversation — it needs a prompt, a voice and no round trip — and that a model small enough to embed can win the places where cloud TTS never reached.

Both bets can be right. The test over the next year is whether a 31M model can hold a multi-turn character without context conditioning, or whether teams end up pairing a small TTS model with a separate state layer that does the job CSM-1B was designed to do internally. Gradium has not published a multi-turn evaluation. Until it does, the safe reading is that Phonon has closed the gap on single-utterance English quality and has not yet shown that it can carry a conversation.