Orpheus-3B Sunbird Multilingual TTS¶

Overview¶

Orpheus-3B Sunbird Multilingual TTS is a multilingual, multi-speaker text-to-speech model fine-tuned from unsloth/orpheus-3b-0.1-pretrained on the full Sunbird/tts corpus, covering 20 language configurations and every speaker present in the dataset. The model accepts arbitrary text and emits 24 kHz mono speech via the SNAC audio codec. Voice selection happens at the prompt level: prepend a chosen speaker_id followed by ": " to the input text, and the model produces audio in that speaker's voice.

Task Type¶

Text-to-Speech (TTS)

Base Architecture¶

Llama-3 architecture (via unsloth/orpheus-3b-0.1-pretrained), paired with the hubertsiuzdak/snac_24khz audio codec (24 kHz, 7 codes per ~12 ms frame) for waveform decoding.

Model Size¶

3B parameters

Quantization Format¶

None (full precision). Weights are bfloat16, with a 16-bit merged LoRA adapter (see Training Details below).

Languages Supported¶

20 languages/configs, spanning multiple African language families. Speaker IDs encode both source corpus (salt_*, waxal_*, slr32_*, slr129_*, bateesa_*) and language:

Some configs (lgg, sot, tsn) are present in the training mix but don't currently expose individual voice IDs in this checkpoint. Speaker IDs can be enumerated directly from the Sunbird/tts dataset (see the model card's discovery snippet).

Intended Use¶

Intended for: - Multilingual voice synthesis for accessibility, language learning, human–computer interaction, audio content creation, and downstream speech research across the 20 covered languages - Reference checkpoint for the Sunbird/tts → Orpheus-3B multilingual fine-tuning pipeline (reproducible recipe in Orpheus_3B_Sunbird_Multilingual.ipynb)

Out of scope: - Voice impersonation or deception of identifiable real persons - High-stakes / safety-critical use (medical, legal, emergency) without human review - Languages outside the 20 covered configs - Code-switching within a single prompt - Cross-language voice transfer (e.g. sending Acholi text to a Luganda speaker_id) — language identity travels via the speaker tag, so text language must match the chosen speaker's language

Known Limitations¶

• Quality varies by language — the training corpus is unbalanced across the 20 configs; languages with more speaker hours produce more natural speech. Audition per-language samples before relying on a voice in production.
• No language conditioning — there is no separate language token; mismatching speaker_id and text language is undefined behaviour.
• Vocabulary coverage limited to each config's training subset; unfamiliar words, code-switching, and out-of-distribution proper nouns may cause artefacts.
• Long utterances may degrade or truncate beyond ~10 seconds of speech (trained on utterances up to ~16s, max_seq_length=4096).
• Sampling variance — with do_sample=True, identical prompts can produce different deliveries between runs; pass seed= for reproducibility.
• No emotion/style control — unlike the upstream orpheus-3b-0.1-ft, this fine-tune was not exposed to in-text emotion tags (e.g. <laugh>, <sigh>); such tags are tokenized as ordinary text with no special effect.
• Inherits any biases present in the Sunbird/tts corpus and Llama-3's pretraining; not systematically audited per language.
• No automated quality metrics (WER, MOS, language-confusion eval) have been run for this release — evaluation was qualitative, on a held-out test sample.

Hardware Requirements¶

Audio decoding via SNAC runs on CPU and adds ~50–150 ms per utterance.

How to Load / Run¶

Two inference paths are supported: transformers + Unsloth (best for development/notebooks) and vLLM (best for high-throughput serving, batched requests, deployment).

Option A — transformers + Unsloth¶

pip install unsloth snac soundfile torchcodec "datasets>=3.4.1,<4.0.0"

import os
import numpy as np
import torch
import soundfile as sf
from unsloth import FastLanguageModel
from snac import SNAC

MODEL_ID = "sunbird/orpheus-3b-tts-multilingual"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=MODEL_ID,
    max_seq_length=4096,
    dtype=None,
    load_in_4bit=False,
    token=os.environ.get("HF_TOKEN"),
)
FastLanguageModel.for_inference(model)

snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to("cpu")

# See full model card for synthesize() implementation and token constants.
wav = synthesize("Mwattu, Mukama yeebazibwe.", speaker_id="salt_lug_0001", seed=42)
sf.write("luganda.wav", wav, 24000)

Option B — vLLM (high throughput / batched)¶

pip install vllm snac soundfile torchcodec "datasets>=3.4.1,<4.0.0"

Note: vLLM ships its own torch/transformers versions and conflicts with Unsloth's pinned versions — use a separate Python environment for vLLM serving.

from vllm import LLM, SamplingParams

llm = LLM(
    model="sunbird/orpheus-3b-tts-multilingual",
    dtype="bfloat16",
    max_model_len=4096,
    gpu_memory_utilization=0.85,
)

# See full model card for _build_prompt_token_ids(), synthesize(), and
# synthesize_batch() implementations, which support mixing different
# speaker_ids in a single batched call.

Generation parameters¶

Training Details¶

Training data was drawn from all 20 configs of Sunbird/tts (train + test splits concatenated, no speaker filter), with audio encoded via hubertsiuzdak/snac_24khz and tagged per-row with speaker_id so the model learns the multi-speaker prompt format across every voice.

Licence¶

Apache 2.0 (matching the upstream unsloth/orpheus-3b-0.1-pretrained license). Transitively inherits attribution obligations from the Orpheus-TTS project (CanopyAI), the Llama-3 base architecture (Meta Llama 3 Community License), the SNAC codec (MIT), and the Sunbird/tts dataset and its voice donors.

Hugging Face¶

huggingface.co/Sunbird/orpheus-3b-tts-multilingual

Related Model¶

A single-speaker variant is also available for those who only need one voice: sunbird/orpheus-3b-tts-salt-lug-0001 (same recipe, scoped to a single Luganda speaker).

Last Updated¶

2026

Citation¶

@misc{sunbird_orpheus3b_multilingual_2026,
  title        = {Orpheus-3B Sunbird Multilingual TTS},
  author       = {Sunbird AI},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/sunbird/orpheus-3b-tts-multilingual}},
}

@misc{sunbird_tts_dataset,
  title        = {Sunbird Speech Dataset},
  author       = {Sunbird AI},
  howpublished = {\url{https://huggingface.co/datasets/Sunbird/tts}},
}

@misc{orpheus_tts_2025,
  title        = {Orpheus-TTS},
  author       = {Canopy Labs},
  year         = {2025},
  howpublished = {\url{https://github.com/canopyai/Orpheus-TTS}},
}