Grok Voice Transcribe 2.0 发布：转写准确率翻倍，价格不变
英文标题：🔶 Grok Voice Transcribe 2.0
Grok 发布语音转文字模型 Grok Voice Transcribe 2.0，称其在真实世界评测中的准确率达到上一代 Grok Voice Transcribe 1.0 的两倍，而价格保持不变。该模型在 Artificial Analysis 公开榜单的 32 个流式模型中准确率排名第一，并针对电话客服、对话、口述账号信息及多语言短语等生产场景进行了优化。在 19 种语言的短语测试集中，其词错误率从 20.6% 降至 6.8%。现有 Speech-to-Text API 集成无需修改代码即可获得准确率提升，但 Grok Voice Transcribe 1.0 将在未来数周内弃用。
576.5 KB
中文
今天，我们发布最新的语音转文字模型 Grok Voice Transcribe 2.0。在真实世界评测中，Grok Voice Transcribe 2.0 是目前准确率最高的转写模型之一；在价格相同的情况下，其准确率是 Grok Voice Transcribe 1.0 的两倍。
Grok Voice Transcribe 2.0 基于支撑 Grok Voice 的音频基础模型构建。Grok Voice 每天已为数万通客户支持电话提供支持，转写数百万小时的视频旁白，并在实体产品中运行语音 Agent，其中包括 Tesla 车辆中的 Grok 助手。该模型使用在多样化环境中录制的真实、嘈杂、多语言音频组成的独特数据集训练，并通过后训练进行优化。
最终成果是：它成为面向真实世界语音场景最准确的转写模型之一。
大多数转写模型在干净的单人语音上表现良好，但真实世界的音频更加复杂：不稳定的电话线路、交叠的人声、本地口音，以及被口述出来的电话号码或电子邮件地址。我们针对各种条件和环境下最具挑战性的音频，构建了 Grok Voice Transcribe 2.0。
在公开的 Artificial Analysis 排行榜上，Grok Voice Transcribe 2.0 在 32 个流式模型中准确率排名第一。
准确率与价格
越靠右上越好。
2%3%4%5%6%7%8%$0$3$6$9$12$15$18词错误率（%） • 更优 →价格（美元 / 1,000 分钟） • 更优 →Grok Voice Transcribe 2.0Grok Voice Transcribe 1.0Muse Voice TranscribeScribe v2 RealtimeGPT Live TranscribeGemini 3.5 Transcribe LiveUniversal-3.5 Realtime ProChirp 3 StreamingAzure STT Real-timeNova-3 RealtimeFlux
除公开基准测试外，我们还在四个来自生产流量的内部数据集上衡量词错误率：客户支持电话的电话音频、与 Grok 的对话、口述凭证信息（如账号代码和电子邮件地址），以及多语言短语音指令。Grok Voice Transcribe 2.0 在这四项测试中均优于 Grok Voice Transcribe 1.0；在电话音频测试中，它领先于我们测试的所有模型。
电话音频（8 kHz）
客户支持电话 · 英语
对话
与 Grok 的对话 · 英语
凭证信息
电话号码、电子邮件地址、住址 · 英语
短语
语音助手指令 · 19 种语言
Grok Voice Transcribe 2.0 可转写数十种语言，能够自动检测语言，并在单次处理过程中跟随录音中的语言切换。多语言准确率是其相对于 Grok Voice Transcribe 1.0 提升最大的部分。
转写准确率
词错误率（%），越低越好。
Grok Voice Transcribe 2.0
Grok Voice Transcribe 1.0
车载指令等短语能为模型识别语言提供的上下文很少。在我们的短语数据集中，词错误率从 20.6% 降至 6.8%。
Grok Voice Transcribe 2.0 支持高级配置与控制功能。现有 Speech-to-Text API 集成无需修改代码，即可获得准确率提升：
- 批量与流式处理。 可转写已录制的文件和 URL，也可实时转写音频流。
- 词级时间戳。 每个词都带有精确的起止时间和置信度评分。
- 说话人分离。 在转写文本中标记每位说话人，无额外费用。
- 多通道转写。 最多可独立转写 8 个通道。
- 关键术语偏置。 每次请求最多可传入 100 个领域术语，例如产品名称或医疗词汇。
- 文本格式化。 数字、日期、货币、电话号码和电子邮件地址将以书面形式返回。
- 填充词移除。 从转写文本中省略“um”和“uh”等填充词。
- 智能轮次检测。 为语音 Agent 检测说话人轮次结束点。
Atlassian Loom 被广泛用于录制和分享屏幕录制内容。Atlassian 发现，Grok Voice Transcribe 2.0 在转写 Loom 视频时比其现有解决方案更准确。准确的转写文本将开启新的 AI 工作流：在 Loom 中录制行动计划，将转写文本导入 Cursor，Cursor 即可直接完成代码更新。
“我们始终相信，推动工作前进的最佳方式是一次性捕捉上下文，并让它流向所有地方。Grok 为 Loom 提供语音转文字能力，Cursor 再将其转化为代码，我们由此闭合了从上下文到代码的循环：录下你的意图，工作便会完成。这让我们得以一窥 AI 辅助开发的发展方向。”
Sanchan Saxena，Atlassian Teamwork Collection 高级副总裁
Grok Voice Transcribe 2.0 的定价与 Grok Voice Transcribe 1.0 完全相同。批量转写仍为每小时音频 0.10 美元，流式转写为每小时 0.20 美元，其中包括说话人分离、时间戳和关键术语功能。
批量处理
每小时音频的美元价格
流式处理
每小时音频的美元价格
Grok Voice Transcribe 2.0 很快将成为 Speech-to-Text API 的默认模型，Grok Voice Transcribe 1.0 将在未来数周内弃用。如需在迁移期间继续使用旧模型，请固定使用 grok-voice-transcribe-1.0。
开始使用 Grok Voice Transcribe
English / 原文
Today we're releasing Grok Voice Transcribe 2.0, our latest speech-to-text model. Across our real-world evaluations, Grok Voice Transcribe 2.0 is one of the most accurate transcription models available today and twice as accurate as Grok Voice Transcribe 1.0, at the same price.
Grok Voice Transcribe 2.0 is built on the audio foundation model behind Grok Voice. Grok Voice already powers tens of thousands of customer-support calls a day, transcribes millions of hours of video narration, and runs voice agents in physical products, including the Grok assistant in Tesla vehicles. It is trained on a unique dataset of live, noisy, multilingual audio recorded across a diverse set of environments and refined with post-training.
The result is one of the most accurate transcription models for speech in real-world settings.
Most transcription models do well on clean, single-speaker audio. Real-world audio is harder: flaky phone lines, competing voices, local accents, and phone numbers or email addresses read aloud. We built Grok Voice Transcribe 2.0 for the hardest audio across conditions and environments.
On the public Artificial Analysis leaderboard, Grok Voice Transcribe 2.0 ranks first for accuracy among 32 streaming models.
Accuracy vs. Price
Upper and to the right is better.
2%3%4%5%6%7%8%$0$3$6$9$12$15$18Word Error Rate (%) • Better →Price ($ / 1,000 min) • Better →Grok Voice Transcribe 2.0Grok Voice Transcribe 1.0Muse Voice TranscribeScribe v2 RealtimeGPT Live TranscribeGemini 3.5 Transcribe LiveUniversal-3.5 Realtime ProChirp 3 StreamingAzure STT Real-timeNova-3 RealtimeFlux
In addition to public benchmarks, we measure word error rate on four internal sets drawn from production traffic: telephony audio from customer-support calls, conversations with Grok, spoken credentials such as account codes and email addresses, and short multilingual voice commands. Grok Voice Transcribe 2.0 improves on Grok Voice Transcribe 1.0 across all four, and on telephony it leads every model we tested.
Telephony (8 kHz)
Customer support calls · English
Conversational
Conversations with Grok · English
Credentials
Phone numbers, emails, addresses · English
Short Phrases
Voice-assistant utterances · 19 languages
Grok Voice Transcribe 2.0 transcribes dozens of languages, detects the language automatically, and follows mid-recording switches in a single pass. Multilingual accuracy is its largest improvement over Grok Voice Transcribe 1.0.
Transcription accuracy
Word Error Rate (%). Lower is better.
Grok Voice Transcribe 2.0
Grok Voice Transcribe 1.0
Short phrases such as in-car commands give the model little context to identify the language. On our short-phrase set, word error rate drops from 20.6% to 6.8%.
Grok Voice Transcribe 2.0 supports advanced configuration and controls. Existing Speech-to-Text API integrations get the accuracy improvement with no code changes:
- Batch and streaming. Transcribe recorded files and URLs, or transcribe an audio stream in real time.
- Word-level timestamps. Each word has precise start/end times and confidence scores.
- Speaker diarization. Label each speaker in the transcript, at no additional cost.
- Multichannel transcription. Transcribe up to 8 channels independently.
- Key term biasing. Pass up to 100 domain terms per request, such as product names or medical vocabulary.
- Text formatting. Numbers, dates, currencies, phone numbers, and email addresses are returned in written form.
- Filler word removal. Omit fillers such as "um" and "uh" from the transcript.
- Smart turn detection. Detect the end of a speaker's turn for voice agents.
Atlassian Loom is widely used for recording and sharing screen recordings. Atlassian found Grok Voice Transcribe 2.0 more accurate than their existing solution for transcribing Loom videos. Accurate transcripts open up new AI workflows: record an action plan in Loom, pipe the transcript into Cursor, and it makes the code updates directly.
“We've always believed the best way to move work forward is to capture context once and let it flow everywhere. With Grok powering Loom's speech-to-text and Cursor turning that into code, we're closing the loop from context to code: record what you mean, and the work gets done. It's a glimpse of where AI-assisted development is headed.”
Sanchan Saxena, SVP of Teamwork Collection, Atlassian
Grok Voice Transcribe 2.0 pricing is identical to Grok Voice Transcribe 1.0. Batch transcription remains $0.10 per hour of audio and streaming $0.20 per hour, with diarization, timestamps, and key terms included.
Batch
USD per hour of audio
Streaming
USD per hour of audio
Grok Voice Transcribe 2.0 will soon be the default in the Speech-to-Text API, and Grok Voice Transcribe 1.0 will be deprecated in the coming weeks. To stay on it during the transition, pin grok-voice-transcribe-1.0.