decibel-hq/reso1-3b-en

Reso-1 3B (English) — expressive, streaming TTS

Reso-1 is Decibel Labs' first voice model: an expressive, tag-controllable, streaming English TTS. LoRA fine-tune of Orpheus 3B (Llama-3 backbone + SNAC 24 kHz codec).

Research preview — private. Weights derive from experimental training data and are not licensed for redistribution or production use.

Samples (novel text, not in training data)

laugh

sigh

price

phone

ask

short

long

convo

whisper

chuckle

Usage

Prompt format: reso: <your text> with inline paralinguistic tags: <laugh> <chuckle> <sigh> <gasp> <whisper> <pause>

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "canopylabs/orpheus-3b-0.1-ft", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "decibel-hq/reso1-3b-en")
tokenizer = AutoTokenizer.from_pretrained("decibel-hq/reso1-3b-en")
# sequence format + SNAC decode: 7 audio tokens/frame from id 128266,
# wrapped in [128259] text [128009,128260,128261,128257] audio [128258,128262]

Sampling: temperature 0.6, top_p 0.95, repetition_penalty >= 1.1 (required — lower values cause audio loops). Streaming: decode SNAC frames incrementally (7 tokens = ~85 ms audio); a GGUF build for llama.cpp is at decibel-hq/reso1-3b-en-GGUF.

Training

• Data: single-voice expressive English, 7,342 clips / 8.87 h after QC (Whisper-WER + speech-rate + tag-event verification), from a 6-category corpus: phonetic (Arctic/Harvard), prose, text-normalization stress set, multi-turn dialogue prosody, paralinguistic tags, edge cases (decibel-hq/reso-prompts-v1)
• LoRA r=64/alpha=64 on all attention+MLP projections, bf16, 3 epochs, lr 2e-4 cosine, effective batch 16, single A100-80GB (~33 min)
• Final train loss 3.47; 198-clip diagnostic heldout, 0 degenerate generations

Limitations

• Single voice, English only (multilingual Indic line planned on an Orpheus-compatible base)
• <whisper>/<pause> are corpus-taught rather than base-native — less robust
• Research preview: no safety alignment; disclose synthetic audio; do not impersonate

Citation

Decibel Labs — Reso voice models (2026). https://github.com/decibel-hq/reso