SpragAI/qwen3-tts-emotion-tags

Qwen3-TTS with Emotion Tags

This finetune provides inline delivery control for the nine Qwen3-TTS CustomVoice presets facilitating emotion controls directly from the transcript.

[Angry] I told them the whole story last night, and now everyone knows what happened.

It covers nine emotion tags across all nine preset voices, with the speaker embeddings left bit-identical to the base model meaning the source voice remains unchanged.

Tags [Angry] [Sad] [Happy] [Fast] [Gentle] [Tired] [Fearful] [Disgusted] [Surprised]

Voices aiden, dylan, eric, ono_anna, ryan, serena, sohee uncle_fu vivian

Usage

pip install qwen-tts

import soundfile as sf
import torch
from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "SpragAI/qwen3-tts-emotion-tags",
    dtype=torch.bfloat16,
    attn_implementation="sdpa",
    device_map="cuda:0",
)

wav, sr = model.generate_custom_voice(
    text="[Sad] None of it ever happened in the end.",
    speaker="ryan", language="Auto")
sf.write("out.wav", wav[0], sr)

Samples

I told them the whole story last night, and now everyone knows what happened.

no tag (reference)

[Angry]

[Sad]

[Happy]

[Fast]

[Gentle]

[Tired]

[Fearful]

[Disgusted]

[Surprised]

Limitations

• Single utterance tags. A tag conditions the whole utterance wherever it appears.
• English only. Other languages may work but were not included in the training corpus.

Training

base	Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
method	LoRA r=32, alpha=64, dropout 0.05
targets	q,k,v,o,gate,up,down_proj (talker body)
also trainable	the 15 code_predictor codec input embeddings
frozen	code_predictor network, codec_embedding (speaker rows preserved)
optimiser	Prodigy, d_coef 1.0, batch 4, 3 epochs
corpus	76862 rows, 9 speakers, 9 tags, 16.7% with an explicit English prefix