whatmodel.sh / tts / Orpheus TTS 3B 0.1 (finetune-prod)

Orpheus TTS 3B 0.1 (finetune-prod)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Orpheus TTS 3B 0.1 (finetune-prod) is an open-weights text-to-speech model from Canopy Labs, released March 2025. Its published size is 3B parameters. Licence: apache-2.0.

Catalogue identity notes

HF checkpoint canopylabs/orpheus-tts-0.1-finetune-prod (English finetuned voices, e.g. tara) as mapped in EmergentTTS evaluation_runner.py. Keep distinct from the unversioned Vapi 'Orpheus' entry and from multilingual research previews.

Model information sources

· Weights licence ↗ · Parameter count ↗

Maker's size label (card at sha 4206a56e5a68cf6cf96900a8a78acd3370c02eb6; GitHub README e64661f: 'built on the Llama-3b backbone'). Judgement: HF's safetensors metadata for this repo reports total 3782986752 parameters (≈3.78B, extended Llama-3.2-3B vocab); that is a Hub-computed count, not a maker statement, so the maker's '3B' is given. Excludes the SNAC decoder.

47.4±4.6#73 of 118 text-to-speech modelsfrom 3 benchmark results across 2 sources

Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.

$ how it compares

The text-to-speech models ranked either side of Orpheus TTS 3B 0.1 (finetune-prod), out of 148 with a score.

#modelscorewhereRAMreleased
90FireRedTTS (v1) FireRedTeam (Xiaohongshu)48±3local~3.1GB2024-09
91Vokan ShoukanLabs48±14comparison only2024-03
92Orpheus TTS 3B 0.1 (finetune-prod) Canopy Labs47±5api $15.00/M chars +local~2.4GB2025-03
93Kyutai TTS 1.6B en/fr Kyutai47±5local~4.1GB2025-07
94MiMo v2.5 TTS Xiaomi MiMo47±4comparison only2026-04

see all text-to-speech models · find the ones that fit your machine

$ what it's good at

other tests3 not in a task · 91% of its score
testresultvs the others in its rundifficultycountswhy
EmergentTTS · 6-category preference29.4% ✎▅▅▃▃█▅▅▃·▃·▅ beats 9/16hard · avg 30%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 56%at its level
EmergentTTS · v1 paper automatic MOS (WV-MOS)3.76 / 5 ✎▄·█·█▄····█▄ beats 5/8easy · avg 3.93 / 5▇▇▇▇▇▇▇▇ 19%blunt · at its level
EmergentTTS · 6-category WER17.7% ✎▂▄··▂▄▄··█▄▆ beats 6/16avg 16%▇▇▇▇▇▇ 16%blunt

$ where to run it

Through an API

api · Together AI · $15.00/M chars · docs.together.ai ↗ · price sourceUSD per 1M input characters: Together's serverless model list gives Canopy Labs Orpheus 3B, API model canopylabs/orpheus-3b-0.1-ft, at $15.00 per 1M chars. Identity: the Hugging Face repo canopylabs/orpheus-tts-0.1-finetune-prod redirects to canopylabs/orpheus-3b-0.1-ft (same repo id), so this is the same checkpoint. Product: Together serverless text-to-speech. API model: canopylabs/orpheus-3b-0.1-ft. Settings: standard. Price basis: published list rate. Exclusions: DeepInfra lists the same checkpoint at $7.00/1M characters (smaller host, not used). Currency: USD. Together official docs, observed 2026-09-25.

Other ways to run it

version

  1. Run it as published · needs ~2.42 GB · from github.com/canopyai
    pip install orpheus-cpp scipy && pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu && python -c 'from scipy.io.wavfile import write from orpheus_cpp import OrpheusCpp import numpy as np orpheus = OrpheusCpp(verbose=False, lang="en") text = "Hello, world." buffer = [] for i, (sr, chunk) in enumerate(orpheus.stream_tts_sync(text, options={"voice_id": "tara"})): buffer.append(chunk) print(f"Generated chunk {i}") buffer = np.concatenate(buffer, axis=1) write("output.wav", 24_000, np.concatenate(buffer))'

Official weights and documentation: huggingface.co/canopylabs ↗

$ how the score was worked out · EmergentTTS-Eval six-category… 72%, EmergentTTS-Eval v1 paper (ar… 19%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
EmergentTTS-Eval six-category leaderboard2▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 72%Apache-2.0
EmergentTTS-Eval v1 paper (arXiv:2505.23009v1) · Table 1 automatic MOS1▇▇▇▇▇▇ 19%CC BY 4.0
prior · pull toward the middle▇▇▇ 9%
weighted average= 47.4

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json