marcosremar2/qwen3-omni-ptbr-qwen3tts-synthetic

Qwen3 Omni PT-BR Synthetic TTS Teacher Dataset

This dataset contains synthetic Portuguese (Brazil) speech/text examples for an experimental omni speech pipeline:

Whisper/ASR -> Tucano hidden states -> Qwen3 audio-code talker -> Qwen3-TTS tokenizer decoder

Contents

• fast_pass_answer_texts.jsonl: 50,000 synthetic Q&A/text rows.
• teacher_wavs_det_v1/: 12,204 generated WAV files.
• audio_metadata.jsonl: lightweight audio manifest with relative WAV paths and text.
• fast_pass_teacher_codes_det_v1.jsonl: teacher manifest including Qwen3-TTS audio codes.
• dataset_summary.json: count and generation summary.

Generation

The text rows were generated synthetically for PT-BR omni-model training. The audio rows were generated with a Qwen3-TTS teacher pipeline:

• Teacher model: Qwen/Qwen3-TTS-12Hz-1.7B-Base
• Tokenizer/codec target: Qwen/Qwen3-TTS-Tokenizer-12Hz
• Voice engine: qwen3_tts_teacher_voice_clone
• Speaker id: venus_iasini
• Sample rate: 24 kHz
• Quantizers: 16

The audio generation target was 50,000 examples, but this snapshot was intentionally stopped at 12,204 generated audio examples.

Intended Use

This is a research/prototype dataset for training and evaluating a compact Qwen3-audio-code talker conditioned on PT-BR LLM hidden states.

Important Notice

The speech files are synthetic outputs generated by Qwen3-TTS. If you use, redistribute, or train from this dataset, ensure you have the necessary rights for the speaker identity/reference material and comply with the licenses and terms of the source models and voice references.