最新模型
Grok Voice Transcribe 2.0 reduces transcription errors; watch the default model switch
The new version retains existing transcription pricing, while both official multilingual short-utterance tests and independent batch benchmarks show lower word error rates. Batch speed, Chinese formatting support, and the default model switch still require separate verification.
SpaceXAI launched Grok Voice Transcribe 2.0 on September 18, retaining batch pricing of $0.10 per hour of audio and streaming pricing of $0.20 per hour. The company says the model was trained on noisy, multilingual audio and refined through post-training to improve recognition for phone calls, dictated account numbers, and short commands. Word error rate on its internal multilingual short-utterance test fell from 20.6% to 6.8%, though this remains a vendor-reported measurement. [Launch announcement](https://x.ai/news/grok-voice-transcribe-2)
Independent evidence comes from Artificial Analysis: when checked on September 19, its non-streaming leaderboard showed a word error rate of 2.3% for 2.0, compared with 4.0% for 1.0. However, the speed multiplier in the same table fell from 242.1 to 155.1, showing that improved accuracy does not mean every workload runs faster. These are batch measurements for complete audio recordings and cannot be directly translated into the time users wait for a voice agent to respond. [Batch benchmark](https://artificialanalysis.ai/speech-to-text)
The API continues to offer file transcription and WebSocket streaming, with optional speaker labels and separate transcription for up to eight audio channels. Smart Turn predicts whether a speaker has finished speaking; number formatting requires both an explicit language setting and `format=true`. For engineering teams, these options affect turn-taking timing, data fields, and subsequent tool calls. After updating the model, teams should replay the same audio to check whether entity values and response timing have changed. [API guide](https://docs.x.ai/developers/model-capabilities/audio/speech-to-text)
Version migration also warrants attention: the launch announcement says 2.0 will soon become the default, but the current model documentation already lists it as the default when `model` is omitted. Services that require reproducible results should explicitly specify a version. The company has announced plans to deprecate 1.0 in the coming weeks, but the announcement provides no exact date. [Model documentation](https://docs.x.ai/developers/models/speech-to-text)
Evidence gaps remain for Chinese deployments. Chinese is absent from the official list of supported formatting languages, so claims of multilingual support cannot establish the output contract for Chinese numbers and addresses. The composition of the independent leaderboard’s roughly eight-hour dataset is also no substitute for testing Traditional Chinese or Chinese-English code-switching. Agent conversations account for half of the benchmark data, with parliamentary speeches and earnings calls each accounting for a quarter; numbers and punctuation are normalized. If a product requires verbatim preservation or exact reconstruction of account-number formatting, leaderboard scores may still differ from actual acceptance criteria. [Benchmark methodology](https://artificialanalysis.ai/methodology/speech-to-text)
Streaming benchmarks measure latency from the endpoint identified by voice activity detection, which differs from a full conversational round trip. Engineers should next measure character error rate, exact-match accuracy for proper nouns, and premature endpointing rate separately, while recording endpointing settings and the sequence of streaming events, before deciding whether to switch.