/Qwen2-Audio-train

The official repo of Qwen2-Audio chat & pretrained large audio language model proposed by Alibaba Cloud.

Primary LanguagePython

中文  ｜   English  



Qwen2-Audio-7B 🤖 | 🤗  ｜ Qwen-Audio-7B-Instruct 🤖 | 🤗  ｜ Demo 🤖 | 🤗 
📑 Paper    |    📑 Blog    |    💬 WeChat (微信)   |    Discord  

We introduce the latest progress of Qwen-Audio, a large-scale audio-language model called Qwen2-Audio, which is capable of accepting various audio signal inputs and performing audio analysis or direct textual responses with regard to speech instructions. We introduce two distinct audio interaction modes:

  • voice chat: users can freely engage in voice interactions with Qwen2-Audio without text input;
  • audio analysis: users could provide audio and text instructions for analysis during the interaction;

We've released two models of the Qwen2-Audio series: Qwen2-Audio-7B and Qwen2-Audio-7B-Instruct.

Architecture

The overview of three-stage training process of Qwen2-Audio.

News and Updates

  • 2024.8.9 🎉 We released the checkpoints of both Qwen2-Audio-7B and Qwen2-Audio-7B-Instruct on ModelScope and Hugging Face.
  • 2024.7.15 🎉 We released the paper of Qwen2-Audio, introducing the relevant model structure, training methods, and model performance. Check our report for details!
  • 2023.11.30 🔥 We released the Qwen-Audio series.

Evaluation

We evaluated the Qwen2-Audio's abilities on 13 standard benchmarks as follows:

TaskDescriptionDatasetSplitMetric
ASRAutomatic Speech RecognitionFleursdev | testWER
Aishell2test
Librispeechdev | test
Common Voicedev | test
S2TTSpeech-to-Text TranslationCoVoST2testBLEU
SERSpeech Emotion RecognitionMeldtestACC
VSCVocal Sound ClassificationVocalSoundtestACC
AIR-Bench
Chat-Benchmark-SpeechFisher
SpokenWOZ
IEMOCAP
Common voice		dev | testGPT-4 Eval
Chat-Benchmark-SoundClothodev | testGPT-4 Eval
Chat-Benchmark-MusicMusicCapsdev | testGPT-4 Eval
Chat-Benchmark-Mixed-AudioCommon voice
AudioCaps
MusicCaps		dev | testGPT-4 Eval

The below is the overal performance:

The details of evaluation are as follows:
(Note: The evaluation results we present are based on the initial model of the original training framework. However, the scores showed some fluctuations after converting the framework to Huggingface. Here, we present our complete evaluation results, starting with the initial model results from the paper.)

TaskDatasetModelPerformance
MetricsResults
ASRLibrispeech
dev-clean | dev-other |
test-clean | test-other		SpeechT5WER 2.1 | 5.5 | 2.4 | 5.8
SpeechNet- | - | 30.7 | -
SLM-FT- | - | 2.6 | 5.0
SALMONN- | - | 2.1 | 4.9
SpeechVerse- | - | 2.1 | 4.4
Qwen-Audio1.8 | 4.0 | 2.0 | 4.2
Qwen2-Audio1.3 | 3.4 | 1.6 | 3.6
Common Voice 15
en | zh | yue | fr		Whisper-large-v3WER 9.3 | 12.8 | 10.9 | 10.8
Qwen2-Audio8.6 | 6.9 | 5.9 | 9.6
Fleurs
zh		Whisper-large-v3WER 7.7
Qwen2-Audio7.5
Aishell2
Mic | iOS | Android		MMSpeech-baseWER 4.5 | 3.9 | 4.0
Paraformer-large- | 2.9 | -
Qwen-Audio3.3 | 3.1 | 3.3
Qwen2-Audio3.0 | 3.0 | 2.9
S2TTCoVoST2
en-de | de-en |
en-zh | zh-en		SALMONNBLEU 18.6 | - | 33.1 | -
SpeechLLaMA- | 27.1 | - | 12.3
BLSP14.1 | - | - | -
Qwen-Audio25.1 | 33.9 | 41.5 | 15.7
Qwen2-Audio29.9 | 35.2 | 45.2 | 24.4
CoVoST2
es-en | fr-en | it-en |		SpeechLLaMABLEU 27.9 | 25.2 | 25.9
Qwen-Audio39.7 | 38.5 | 36.0
Qwen2-Audio40.0 | 38.5 | 36.3
SERMeldWavLM-largeACC 0.542
Qwen-Audio0.557
Qwen2-Audio0.553
VSCVocalSoundCLAPACC 0.4945
Pengi0.6035
Qwen-Audio0.9289
Qwen2-Audio0.9392
AIR-Bench
Chat Benchmark
Speech | Sound |
Music | Mixed-Audio		SALMONN
BLSP
Pandagpt
Macaw-LLM
SpeechGPT
Next-gpt
Qwen-Audio
Gemini-1.5-pro
Qwen2-Audio		GPT-4 6.16 | 6.28 | 5.95 | 6.08
6.17 | 5.55 | 5.08 | 5.33
3.58 | 5.46 | 5.06 | 4.25
0.97 | 1.01 | 0.91 | 1.01
1.57 | 0.95 | 0.95 | 4.13
3.86 | 4.76 | 4.18 | 4.13
6.47 | 6.95 | 5.52 | 6.08
6.97 | 5.49 | 5.06 | 5.27
7.18 | 6.99 | 6.79 | 6.77

(Second is after converting huggingface)

TaskDatasetModelPerformance
MetricsResults
ASRLibrispeech
dev-clean | dev-other |
test-clean | test-other		SpeechT5WER 2.1 | 5.5 | 2.4 | 5.8
SpeechNet- | - | 30.7 | -
SLM-FT- | - | 2.6 | 5.0
SALMONN- | - | 2.1 | 4.9
SpeechVerse- | - | 2.1 | 4.4
Qwen-Audio1.8 | 4.0 | 2.0 | 4.2
Qwen2-Audio1.7 | 3.6 | 1.7 | 4.0
Common Voice 15
en | zh | yue | fr		Whisper-large-v3WER 9.3 | 12.8 | 10.9 | 10.8
Qwen2-Audio8.7 | 6.5 | 5.9 | 9.6
Fleurs
zh		Whisper-large-v3WER 7.7
Qwen2-Audio7.0
Aishell2
Mic | iOS | Android		MMSpeech-baseWER 4.5 | 3.9 | 4.0
Paraformer-large- | 2.9 | -
Qwen-Audio3.3 | 3.1 | 3.3
Qwen2-Audio3.2 | 3.1 | 2.9
S2TTCoVoST2
en-de | de-en |
en-zh | zh-en		SALMONNBLEU 18.6 | - | 33.1 | -
SpeechLLaMA- | 27.1 | - | 12.3
BLSP14.1 | - | - | -
Qwen-Audio25.1 | 33.9 | 41.5 | 15.7
Qwen2-Audio29.6 | 33.6 | 45.6 | 24.0
CoVoST2
es-en | fr-en | it-en |		SpeechLLaMABLEU 27.9 | 25.2 | 25.9
Qwen-Audio39.7 | 38.5 | 36.0
Qwen2-Audio38.7 | 37.2 | 35.2
SERMeldWavLM-largeACC 0.542
Qwen-Audio0.557
Qwen2-Audio0.535
VSCVocalSoundCLAPACC 0.4945
Pengi0.6035
Qwen-Audio0.9289
Qwen2-Audio0.9395
AIR-Bench
Chat Benchmark
Speech | Sound |
Music | Mixed-Audio		SALMONN
BLSP
Pandagpt
Macaw-LLM
SpeechGPT
Next-gpt
Qwen-Audio
Gemini-1.5-pro
Qwen2-Audio		GPT-4 6.16 | 6.28 | 5.95 | 6.08
6.17 | 5.55 | 5.08 | 5.33
3.58 | 5.46 | 5.06 | 4.25
0.97 | 1.01 | 0.91 | 1.01
1.57 | 0.95 | 0.95 | 4.13
3.86 | 4.76 | 4.18 | 4.13
6.47 | 6.95 | 5.52 | 6.08
6.97 | 5.49 | 5.06 | 5.27
7.24 | 6.83 | 6.73 | 6.42

We have provided all evaluation scripts to reproduce our results. Please refer to eval_audio/EVALUATION.md for details.

Requirements

The code of Qwen2-Audio has been in the latest Hugging face transformers and we advise you to build from source with command pip install git+https://github.com/huggingface/transformers, or you might encounter the following error:

KeyError: 'qwen2-audio'

Quickstart

Below, we provide simple examples to show how to use Qwen2-Audio and Qwen2-Audio-Instruct with 🤗 Transformers. Before running the code, make sure you have setup the environment and installed the required packages. Make sure you meet the above requirements, and then install the dependent libraries. Now you can start with ModelScope or Transformers. Qwen2-Audio models currently perform best with audio clips under 30 seconds.

🤗 Transformers

In the following, we demonstrate how to use Qwen2-Audio-7B-Instruct for the inference, supporting both voice chat and audio analysis modes. Note that we have used the ChatML format for dialog, in this demo we show how to leverage apply_chat_template for this purpose.

Voice Chat Inference

In the voice chat mode, users can freely engage in voice interactions with Qwen2-Audio without text input:

from io import BytesIO
from urllib.request import urlopen
import librosa
from transformers import Qwen2AudioForConditionalGeneration, AutoProcessor

processor = AutoProcessor.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct")
model = Qwen2AudioForConditionalGeneration.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct", device_map="auto")

conversation = [
    {"role": "user", "content": [
        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/guess_age_gender.wav"},
    ]},
    {"role": "assistant", "content": "Yes, the speaker is female and in her twenties."},
    {"role": "user", "content": [
        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/translate_to_chinese.wav"},
    ]},
]
text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
audios = []
for message in conversation:
    if isinstance(message["content"], list):
        for ele in message["content"]:
            if ele["type"] == "audio":
                audios.append(librosa.load(
                    BytesIO(urlopen(ele['audio_url']).read()), 
                    sr=processor.feature_extractor.sampling_rate)[0]
                )

inputs = processor(text=text, audios=audios, return_tensors="pt", padding=True)
inputs.input_ids = inputs.input_ids.to("cuda")

generate_ids = model.generate(**inputs, max_length=256)
generate_ids = generate_ids[:, inputs.input_ids.size(1):]

response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
Audio Analysis Inference

In the audio analysis, users could provide both audio and text instructions for analysis:

from io import BytesIO
from urllib.request import urlopen
import librosa
from transformers import Qwen2AudioForConditionalGeneration, AutoProcessor

processor = AutoProcessor.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct")
model = Qwen2AudioForConditionalGeneration.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct", device_map="auto")

conversation = [
    {'role': 'system', 'content': 'You are a helpful assistant.'}, 
    {"role": "user", "content": [
        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/glass-breaking-151256.mp3"},
        {"type": "text", "text": "What's that sound?"},
    ]},
    {"role": "assistant", "content": "It is the sound of glass shattering."},
    {"role": "user", "content": [
        {"type": "text", "text": "What can you do when you hear that?"},
    ]},
    {"role": "assistant", "content": "Stay alert and cautious, and check if anyone is hurt or if there is any damage to property."},
    {"role": "user", "content": [
        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/1272-128104-0000.flac"},
        {"type": "text", "text": "What does the person say?"},
    ]},
]
text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
audios = []
for message in conversation:
    if isinstance(message["content"], list):
        for ele in message["content"]:
            if ele["type"] == "audio":
                audios.append(
                    librosa.load(
                        BytesIO(urlopen(ele['audio_url']).read()), 
                        sr=processor.feature_extractor.sampling_rate)[0]
                )

inputs = processor(text=text, audios=audios, return_tensors="pt", padding=True)
inputs.input_ids = inputs.input_ids.to("cuda")

generate_ids = model.generate(**inputs, max_length=256)
generate_ids = generate_ids[:, inputs.input_ids.size(1):]

response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
Batch Inference

We also support batch inference:

from io import BytesIO
from urllib.request import urlopen
import librosa
from transformers import Qwen2AudioForConditionalGeneration, AutoProcessor

processor = AutoProcessor.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct")
model = Qwen2AudioForConditionalGeneration.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct", device_map="auto")

conversation1 = [
    {"role": "user", "content": [
        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/glass-breaking-151256.mp3"},
        {"type": "text", "text": "What's that sound?"},
    ]},
    {"role": "assistant", "content": "It is the sound of glass shattering."},
    {"role": "user", "content": [
        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/f2641_0_throatclearing.wav"},
        {"type": "text", "text": "What can you hear?"},
    ]}
]

conversation2 = [
    {"role": "user", "content": [
        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/1272-128104-0000.flac"},
        {"type": "text", "text": "What does the person say?"},
    ]},
]

conversations = [conversation1, conversation2]

text = [processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False) for conversation in conversations]

audios = []
for conversation in conversations:
    for message in conversation:
        if isinstance(message["content"], list):
            for ele in message["content"]:
                if ele["type"] == "audio":
                    audios.append(
                        librosa.load(
                            BytesIO(urlopen(ele['audio_url']).read()), 
                            sr=processor.feature_extractor.sampling_rate)[0]
                    )

inputs = processor(text=text, audios=audios, return_tensors="pt", padding=True)
inputs['input_ids'] = inputs['input_ids'].to("cuda")
inputs.input_ids = inputs.input_ids.to("cuda")

generate_ids = model.generate(**inputs, max_length=256)
generate_ids = generate_ids[:, inputs.input_ids.size(1):]

response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)

Running Qwen2-Audio pretrained base model is also simple.

from io import BytesIO
from urllib.request import urlopen
import librosa
from transformers import AutoProcessor, Qwen2AudioForConditionalGeneration

model = Qwen2AudioForConditionalGeneration.from_pretrained("Qwen/Qwen2-Audio-7B" ,trust_remote_code=True)
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-Audio-7B" ,trust_remote_code=True)

prompt = "<|audio_bos|><|AUDIO|><|audio_eos|>Generate the caption in English:"
url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Audio/glass-breaking-151256.mp3"
audio, sr = librosa.load(BytesIO(urlopen(url).read()), sr=processor.feature_extractor.sampling_rate)
inputs = processor(text=prompt, audios=audio, return_tensors="pt")

generated_ids = model.generate(**inputs, max_length=256)
generated_ids = generated_ids[:, inputs.input_ids.size(1):]
response = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]

Finetuning

We would like to thank the Hugging Face open-source community for their contributions, which have made it easy for us to implement model fine-tuning with Accelerate and DeepSpeed. We support both LoRA (Low-Rank Adaptation) and full-parameter fine-tuning, with the code provided by Xiaoming Liu.

cd finetune && bash run.sh

🤖 ModelScope

We strongly advise users especially those in mainland China to use ModelScope. snapshot_download can help you solve issues concerning downloading checkpoints.

Demo

Web UI

We provide code for users to build a web UI demo. Before you start, make sure you install the following packages:

pip install -r requirements_web_demo.txt

Then run the command below and click on the generated link:

python demo/web_demo_audio.py

demos

More impressive cases will be updated on our blog at Qwen's blog.

We Are Hiring

If you are interested in joining us as full-time or intern, please contact us at qwen_audio@list.alibaba-inc.com.

License Agreement

Check the license of each model inside its HF repo. It is NOT necessary for you to submit a request for commercial usage.

Citation

If you find our paper and code useful in your research, please consider giving a star ⭐ and citation 📝 :)

@article{Qwen-Audio,
  title={Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models},
  author={Chu, Yunfei and Xu, Jin and Zhou, Xiaohuan and Yang, Qian and Zhang, Shiliang and Yan, Zhijie  and Zhou, Chang and Zhou, Jingren},
  journal={arXiv preprint arXiv:2311.07919},
  year={2023}
}
@article{Qwen2-Audio,
  title={Qwen2-Audio Technical Report},
  author={Chu, Yunfei and Xu, Jin and Yang, Qian and Wei, Haojie and Wei, Xipin and Guo,  Zhifang and Leng, Yichong and Lv, Yuanjun and He, Jinzheng and Lin, Junyang and Zhou, Chang and Zhou, Jingren},
  journal={arXiv preprint arXiv:2407.10759},
  year={2024}
}

Contact Us

If you are interested to leave a message to either our research team or product team, feel free to send an email to qianwen_opensource@alibabacloud.com.