Introduction

• Qwen-Audio uses over 30 tasks, 8 languages, and various types of audio.

• Qwen-Audio contains an audio encoder and a large language model.

  • Audio Encoder - Whisper Large v2 Encoder.

  • LLM - Qwen 7B.

Audio Encoder

• Initialize from whisper large v2 encoder with 640M parameters.

• Preprocessing:

  • Resample to 16kHz.

  • Converts into 80-channel melspectrogram.

  • Window size of 25ms and a hop size of 10ms.

  • SpecAugment is applied.

Language Model

• Pretrained weights from Qwen-7B.

  • 32-layer transformer decoder.

  • Hidden size of 4096.

  • 7.7B parameters.

     

Multitask Pretraining

• Similar tasks can benefit from knowledge sharing during co-training.

• Tasks that rely on lower-level perceptual abilities can assist tasks that require higher-level understanding or reasoning capabilities.

• Simply mixing datasets will introduces interference.

• Whisper specific tasks and condition info as input special tokens.

  • e.g. VAD, lang id, and timestamp tags.

Multi-task Training Format Framework

• Transcription Tag:

  • <|startoftranscripts|> for speech recognition and speech translation tasks.

  • <|startofanalysis|> tag is utilized for other tasks.

• ​Audio Language Tag:

  • <|en|>, <|zh|>, <|de|>, <|es|>, <|ko|>, <|fr|>, <|ja|>, <|it|>.

  • <|unknown|>  for non-speech, such as natural sounds and music.

Multi-task Training Format Framework

• Task Tag:

  • <|transcribe|>, <|translate|>, <|caption|>, <|analysis|>, <|question-answer|>.

  • Append the corresponding questions after the tag for QA task.

• Text Language Tag: the language of output text sequences.

• Timestamps Tag: <|timestamps|>, <|notimestamps|>

  • Fine-grained word-level timestamp (SRWT).

• Output Instruction: provide output instruction for different subtasks.

Supervised Finetuning

• Manually create demonstrations consist of raw text labels, questions, and answers for each task.

• Utilize GPT-3.5 to generate further questions and answers based on the provided raw text labels.

• Label different audios with "Audio id:" to handle multiple audio inputs.

• Each statement is marked with <im_start> and <im_end>.

• Include pure text instruction data.

Chat Template