phoonnx-orpheus — Orpheus 3B (English), ONNX

ONNX weights for Orpheus TTS by Canopy Labs, packaged for phoonnx. This repository holds converted weights only — no new training was done. Original model and code are Apache-2.0; all credit for the model belongs to Canopy Labs.

Orpheus is a Llama-3.2-3B causal LM whose vocabulary carries 28 672 audio tokens. It emits a flat token stream; every seven tokens form one SNAC frame, which SNAC's decoder turns into 2048 samples at 24 kHz.

Read this before you use it: Orpheus is not a CPU model

A 3B backbone needs about 0.37 s per decode step on 12 CPU cores, and SNAC needs ~82 tokens for every second of audio. That is roughly 37-41x slower than real time on CPU. Canopy Labs announced 1B / 400M / 150M tiers but never released them — their own loader still raises "not supported ... will be released very soon" — so there is no smaller Orpheus to fall back to.

Use this on a GPU. It is mirrored for completeness, not as an offline CPU default.

Files (`orpheus-3b-en-onnx/`)

File	What it is
`model.onnx` + `model.onnx_data`..`model.onnx_data_6`	Llama backbone, fp32, KV-cached. The graph needs all seven `.onnx_data` shards next to it.
`snac_decoder.onnx`	SNAC 24 kHz decoder, fp32, copied from onnx-community/snac_24khz-ONNX (Apache-2.0), unmodified.
`tokenizer.json`	The checkpoint's own BPE, copied from upstream.

model.onnx comes from onnx-community/orpheus-3b-0.1-ft-ONNX, re-verified against the torch reference before mirroring (see below).

Voices

tara, leah, jess, leo, dan, mia, zac, zoe — in Canopy Labs' own order of conversational realism. The voice is a name written into the prompt text, not an embedding or a speaker id.

Emotive tags are ordinary text the same BPE encodes: <laugh>, <chuckle>, <sigh>, <cough>, <sniffle>, <groan>, <yawn>, <gasp>.

Verification

Measured against the torch reference (unsloth/orpheus-3b-0.1-ft, an ungated copy of the gated original) on a 19-token served prompt, fp32 both sides:

Variant	Size	Prefill max abs logit diff	Greedy agreement, 25 steps	ms / decode step	Verdict
`model` (fp32)	12.7 GB	0.166	25/25	369	mirrored
`model_q4`	2.4 GB	8.65	23/25	371	rejected — loses parity and is not faster
`model_q4f16`	2.1 GB	8.53	10/25	96	rejected — fails parity

Only fp32 is mirrored. The quantized variants either lose the greedy path or, in the q4 case, buy nothing: its weights are int4 but its activations stay fp32, so onnxruntime dequantizes on the fly and the step cost is unchanged.

SNAC's decoder is stochastic — it contains a noise block, so two decodes of the same codes differ. Its ONNX export was therefore checked against the model's own run-to-run spread (mean of 40 decodes, relative RMS) rather than by exact diff:

SNAC decoder	Relative RMS vs torch mean	Ratio to noise floor	Verdict
`decoder_model` (fp32)	0.0200	0.89x	pass
`decoder_model_fp16`	0.0233	1.03x	pass
`decoder_model_int8`	0.1356	6.02x	fail
`decoder_model_uint8` / `_quantized`	0.1089	4.83x	fail

decoder_model_q4 and decoder_model_bnb4 are byte-for-byte the same size as fp32 and produce identical output — the quantizer skipped this convolutional graph entirely, so those names are misleading rather than smaller.