基于OpenAI的Whisper构建的高效语音识别模型:faster-whisper

1 faster-whisper介绍

faster-whisper是基于OpenAI的Whisper模型的高效实现,它利用CTranslate2,一个专为Transformer模型设计的快速推理引擎。这种实现不仅提高了语音识别的速度,还优化了内存使用效率。faster-whisper的核心优势在于其能够在保持原有模型准确度的同时,大幅提升处理速度,这使得它在处理大规模语音数据时更加高效。

项目地址:https://github.com/SYSTRAN/faster-whisper

1.1 whisper

OpenAI 的开源模型 whisper,可以执行 99 种语言的语音识别和文字转写。但是 whisper 模型占用计算资源多,命令行使用门槛高。whisper 模型本身还存在一些问题,例如 模型幻听问题(大部分的类似于不断重复同一句话、无语音部分复读莫名内容等都是由于这个原因造成的)。要更好使用 whsiper 模型就需要能够准确调试模型参数。但 whisper 模型参数众多,且命令行使用对使用者有一定要求,而且只有 torch 版可以做到调整参数。使用 VAD 类工具也需要一定的动手能力。

1.2 faster-whisper

faster-whisper是具有完全的 whsiper 模型参数,且自带 VAD加持的 whisper 版本,该版本使用了 CTranslate2 来重新实现 whsiper 模型,CT2 对 transformer 类网络进行了优化,使模型推理效率更高。 相比于 openai/whisper,该实现在相同准确性下速度提高了 4 倍以上,同时使用的内存更少。

所谓 VAD 即 Voice Activity Detection —— 声音活动检测,在语音信号处理中,例如语音增强,语音识别等领域有着非常重要的作用。它的作用是从一段语音(纯净或带噪)信号中标识出语音片段与非语音片段。在语音转写任务中,可以提前将语音和非语音部分分离出来,从而提升 whisper 网络识别速度,并减少模型幻听。

VAD地址:https://github.com/snakers4/silero-vad

1.3 性能对比

在性能方面,faster-whisper展现了显著的优势。例如,在使用Large-v2模型和GPU进行13分钟音频的转录测试中,faster-whisper仅需54秒,而原始Whisper模型需要4分30秒。这一显著的性能提升,意味着在实际应用中,faster-whisper能够更快地处理大量数据,特别是在需要实时或近实时语音识别的场景中。

1.4 技术优势

faster-whisper的技术优势不仅体现在速度上。它还支持8位量化,这一技术可以在不牺牲太多准确度的情况下,进一步减少模型在CPU和GPU上的内存占用。这使得faster-whisper在资源受限的环境中也能高效运行,如在移动设备或嵌入式系统上。

aster-whisper适用于多种场景,特别是那些需要快速、准确的语音识别的应用。例如,在客户服务中,它可以用于实时语音转文字,提高响应速度和服务质量。在医疗领域,faster-whisper可以辅助医生快速转录病历,提高工作效率。此外,它还适用于实时会议记录、多语言翻译、教育辅助等多个领域。

2 faster-whisper安装及使用

2.1 conda环境下安装

pip install faster-whisper

2.2 模型下载

large-v3模型:https://huggingface.co/Systran/faster-whisper-large-v3/tree/main
large-v2模型:https://huggingface.co/guillaumekln/faster-whisper-large-v2/tree/main
large-v2模型:https://huggingface.co/guillaumekln/faster-whisper-large-v1/tree/main
medium模型:https://huggingface.co/guillaumekln/faster-whisper-medium/tree/main
small模型:https://huggingface.co/guillaumekln/faster-whisper-small/tree/main
base模型:https://huggingface.co/guillaumekln/faster-whisper-base/tree/main
tiny模型:https://huggingface.co/guillaumekln/faster-whisper-tiny/tree/main

国内模型地址:

https://aifasthub.com/models/guillaumekln

2.3 使用fater-whisper进行中文语音识别

from faster_whisper import WhisperModel

model_size = "large-v3"

# Run on GPU with FP16
model = WhisperModel('../model/faster-whisper-large-v3', device="cuda", compute_type="float16")

# or run on GPU with INT8
# model = WhisperModel(model_size, device="cuda", compute_type="int8_float16")
# or run on CPU with INT8
# model = WhisperModel(model_size, device="cpu", compute_type="int8")

segments, info = model.transcribe("../data/gusuiyizhi.wav", beam_size=5)

print("Detected language '%s' with probability %f" % (info.language, info.language_probability))

for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

参数说明:

  • local_files_only=True 表示加载本地模型
  • model_size_or_path=path 指定加载模型路径
  • device=”cuda” 指定使用cuda
  • compute_type=”int8_float16″ 量化为8位
  • language=”zh” 指定音频语言
  • vad_filter=True 开启vad
  • vad_parameters=dict(min_silence_duration_ms=1000) 设置vad参数
     

运行结果显示:

Detected language 'zh' with probability 0.998535
[0.00s -> 2.48s] 骨髓抑制的定义、危害和预防
[2.48s -> 4.48s] 骨髓抑制是一种病症
[4.48s -> 8.14s] 其主要特点是骨髓中血细胞生成减少
[8.14s -> 11.62s] 这种减少会影响所有类型的血细胞
[11.62s -> 14.74s] 包括红细胞、白细胞和血小板
[14.74s -> 15.74s] 通常
[15.74s -> 20.16s] 骨髓抑制是癌症治疗中使用的化疗或放疗的副作用
[20.16s -> 23.42s] 因为这些疗法在杀伤癌细胞的同时
[23.42s -> 26.22s] 也会损害骨髓中的健康细胞
[26.22s -> 29.26s] 骨髓抑制可能会带来如下危害
[29.26s -> 30.86s] 一、贫血
[30.86s -> 33.74s] 骨髓抑制导致红细胞生成减少
[33.74s -> 35.64s] 可能会引起贫血
[35.64s -> 39.94s] 表现为疲劳、虚弱、呼吸急促和面色苍白
[39.94s -> 44.58s] 严重的贫血可能导致心脏问题和日常活动能力下降
[44.58s -> 46.92s] 二、中性粒细胞减少
[46.92s -> 49.52s] 骨髓抑制导致白细胞减少
[49.52s -> 51.96s] 尤其是中性粒细胞减少
[51.96s -> 55.94s] 患者更容易受到细菌、病毒和真菌感染
[56.22s -> 58.48s] 某些情况下可能会很严重
[58.48s -> 59.24s] 并危及生长
[59.24s -> 61.68s] 三、血小板减少
[61.68s -> 64.60s] 骨髓抑制导致血小板生成减少
[64.60s -> 67.02s] 增加出血和淤伤的风险
[67.02s -> 71.10s] 患者可能会流鼻血、牙龈出血、容易淤伤
[71.10s -> 74.98s] 严重的血小板减少症可导致自发性出血
[74.98s -> 78.18s] 小伤口出血时间延长和内出血
[78.18s -> 79.72s] 这可能危及生命
[79.72s -> 81.58s] 四、治疗延误
[81.58s -> 84.06s] 如果患者的血细胞技术太低
[84.06s -> 87.24s] 则可能需要延迟化疗周期或放疗
[87.24s -> 88.82s] 以使骨髓功能恢复
[88.82s -> 89.22s] 三、血小板减少
[89.24s -> 90.46s] 如果延迟太久
[90.46s -> 92.38s] 可能会延误癌症的治疗
[92.38s -> 94.38s] 造成危险的后果
[94.38s -> 96.24s] 五、剂量减少
[96.24s -> 100.06s] 骨髓抑制可能需要减少化疗或放疗的剂量
[100.06s -> 103.56s] 这可能会影响癌症治疗的整体有效性
[103.56s -> 105.88s] 六、延长住院时间
[105.88s -> 109.96s] 严重的感染或出血可能需要住院治疗和监测
[109.96s -> 114.16s] 增加了患者癌症治疗的总体时间和经济成本
[114.16s -> 116.44s] 七、降低生活质量
[116.44s -> 119.22s] 疲劳、感染风险增加和出血病毒
[119.24s -> 123.72s] 病发症会对患者的整体健康和日常生活功能产生负面影响
[123.72s -> 127.84s] 预防肿瘤放化疗引起的骨髓抑制虽然并不容易
[127.84s -> 132.42s] 但有一些方法可以帮助减少风险或减轻严重程度
[132.42s -> 134.02s] 一、精确剂量
[134.02s -> 137.20s] 医生会根据您的身体状况和肿瘤大小
[137.20s -> 139.68s] 来选择最佳的化疗和放疗剂量
[139.68s -> 141.70s] 以减少对骨髓的影响
[141.70s -> 143.60s] 二、药物调整
[143.60s -> 144.48s] 如果可能
[144.48s -> 148.06s] 医生会选择具有较低骨髓抑制风险的药物
[148.18s -> 149.02s] 或调整
[149.24s -> 150.72s] 治疗方案来降低风险
[150.72s -> 152.60s] 三、治疗间隔
[152.60s -> 156.16s] 医生可能会增加治疗之间的间隔时间
[156.16s -> 159.04s] 让您的身体有更多时间恢复
[159.04s -> 160.98s] 四、生长因子
[160.98s -> 163.32s] 医生可能会开出生长因子
[163.32s -> 165.36s] 如粒细胞刺激因子
[165.36s -> 167.88s] GCSF或红细胞生成素
[167.88s -> 168.70s] EPO
[168.70s -> 172.48s] 以帮助身体加速产生白细胞和红细胞
[172.48s -> 174.36s] 五、营养支持
[174.36s -> 176.30s] 保持良好的营养状态
[176.30s -> 179.14s] 可以帮助身体应对治疗带来的副作用
[179.14s -> 179.22s] 三、医生可能会增加治疗之间的间隔时间
[179.24s -> 181.38s] 注意摄入充足的蛋白质
[181.38s -> 183.06s] 维生素和矿物质
[183.06s -> 185.08s] 以支持骨髓功能
[185.08s -> 187.20s] 六、免疫增强剂
[187.20s -> 188.82s] 在某些情况下
[188.82s -> 191.98s] 医生可能会推荐使用免疫增强剂
[191.98s -> 194.66s] 以提高您的免疫系统抵抗力
[194.66s -> 196.42s] 需要您注意的是
[196.42s -> 199.10s] 以上建议可能因个人情况而异
[199.10s -> 202.44s] 请务必与您的医生保持紧密沟通
[202.44s -> 206.34s] 制定最适合您的个性化的防治策略
[206.34s -> 209.14s] 以应对放化疗可能引起的骨髓抑制
[209.14s -> 209.44s] 请务必与您的医生保持紧密沟通

2.4 使用fater-whisper将中文翻译成英文输出

from faster_whisper import WhisperModel

model_size = "large-v3"

# Run on GPU with FP16
model = WhisperModel('../model/faster-whisper-large-v3', device="cuda", compute_type="float16")

# or run on GPU with INT8
# model = WhisperModel(model_size, device="cuda", compute_type="int8_float16")
# or run on CPU with INT8
# model = WhisperModel(model_size, device="cpu", compute_type="int8")

segments, info = model.transcribe("../data/gusuiyizhi.wav", beam_size=5, language='en')
print("Detected language '%s' with probability %f" % (info.language, info.language_probability))

for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

运行结果显示:

Detected language 'en' with probability 1.000000
[0.00s -> 2.46s]  The definition of bone marrow disease, harm and prevention
[2.46s -> 4.32s]  Bone marrow disease is a disease
[4.32s -> 8.06s]  Its main feature is that blood cells in the bone marrow are reduced
[8.06s -> 11.60s]  This reduction will affect all types of blood cells
[11.60s -> 14.56s]  Including red cells, white cells and blood cells
[14.56s -> 20.08s]  Usually, bone marrow disease is a side effect of chemotherapy or chemotherapy used in cancer treatment
[20.08s -> 23.44s]  Because these treatments kill cancer cells at the same time
[23.44s -> 26.20s]  It will also harm the healthy cells in the bone marrow
[26.20s -> 29.24s]  Bone marrow disease may bring the following harm
[29.24s -> 30.86s]  1. Pneumonia
[30.86s -> 33.68s]  Bone marrow disease may cause red cells to decrease
[33.68s -> 35.60s]  It may cause pneumonia
[35.60s -> 39.94s]  It is a sign of fatigue, weakness, respiratory cramps and pale face
[39.94s -> 44.52s]  Severe pneumonia may lead to heart problems and reduced daily activity
[44.52s -> 46.90s]  2. Reducing neutral blood cells
[46.90s -> 49.48s]  Bone marrow disease may lead to reduced white cells
[49.48s -> 51.94s]  Especially reduced neutral blood cells
[51.94s -> 55.52s]  Patients are more likely to be infected with bacteria, viruses and germs
[55.52s -> 59.02s]  In some cases, it may be very serious and critical
[59.02s -> 59.22s]  In some cases, it may be very serious and critical
[59.22s -> 59.70s]  In some cases, it may be very serious and critical
[59.70s -> 61.68s]  3. Reducing neutral blood cells
[61.68s -> 64.54s]  Bone marrow disease may cause reduced blood cells to be generated
[64.54s -> 66.98s]  It may increase the risk of bleeding and bruising
[66.98s -> 71.10s]  Patients may have nosebleeds, gums bleeding, and easy bruising
[71.10s -> 74.98s]  Serious neutral blood cells can lead to auto-blood bleeding
[74.98s -> 78.14s]  Small wound bleeding time is prolonged and internal bleeding
[78.14s -> 79.72s]  This may be fatal to life
[79.72s -> 81.54s]  4. Treating delay
[81.54s -> 84.06s]  If the patient's blood cell technology is too low
[84.06s -> 88.50s]  It may need to delay the chemotherapy cycle or chemotherapy to restore bone marrow function
[88.50s -> 88.94s]  5. Reducing neutral blood cells
[89.22s -> 89.76s]  If the patient's blood cell technology is too low
[89.76s -> 90.60s]  If the patient's blood cell technology is too low
[90.60s -> 92.68s]  It may hinder the treatment of cancer
[92.68s -> 94.50s]  Cause danger
[94.50s -> 97.30s]  5. Reducing the amount of blood
[97.30s -> 101.14s]  Bone marrow disease may require the reduction of chemotherapy or chemotherapy
[101.14s -> 103.46s]  It may affect the overall effectiveness of cancer treatment
[103.46s -> 105.50s]  6. Extending the hospitalization period
[105.50s -> 108.74s]  Severe infection or bleeding may require hospitalization and monitoring
[108.74s -> 113.02s]  It may increase the total time and economic cost of cancer treatment for the patient
[113.02s -> 114.34s]  7. Reducing the quality of life
[114.34s -> 117.06s]  7. Reducing the quality of life
[117.06s -> 118.12s]  3. Reducing the quality of life
[118.12s -> 118.30s]  4. Reduction of life and health

版权声明:本文为博主作者:智慧医疗探索者原创文章,版权归属原作者,如果侵权,请联系我们删除!

原文链接:https://blog.csdn.net/lsb2002/article/details/135286666

共计人评分,平均

到目前为止还没有投票!成为第一位评论此文章。

(0)
xiaoxingxing的头像xiaoxingxing管理团队
上一篇 2024年1月11日
下一篇 2024年1月11日

相关推荐