語音 AI
Azure Speech LLM 2607 Automatically Updates Speech Recognition, Phrase Lists Can Hint More Than 2,000 Entities
Microsoft has updated its managed LLM Speech offering, emphasizing recognition of mixed-language speech, proper names, and numbers, while claiming up to a threefold latency reduction versus version 2605 in some scenarios. The new model uses the existing Fast and real-time APIs but is deployed automatically on the server side, so teams operating existing voice agents must establish their own regression tests and version monitoring.
Microsoft has released Azure AI Speech LLM 2607. Rather than introducing a new set of endpoints, the update replaces the model used by the existing Fast Transcription and real-time speech recognition services. According to Microsoft, the new version improves recognition of mixed-language speech with mid-conversation language switching, Tier 2/Tier 3 locales, punctuation, capitalization, numbers, personal names, and domain-specific entities. Compared with version 2605, processing latency was reduced by as much as threefold in certain internal tests. This is particularly important for voice agents because recognition latency directly consumes the time budget between when a user begins speaking and when the agent starts reasoning.
The most concrete change to the developer interface places hints for proprietary terms in a dedicated `phraseList.phrases` field instead of mixing them into a general-purpose prompt. The list can contain more than 2,000 names, abbreviations, or technical terms. With the Fast API version `2025-10-15`, developers can send an audio file together with a JSON `definition` to `transcriptions:transcribe`. The same configuration can also combine `locales`, speaker diarization, channel selection, and profanity handling. This structured prompting is easier to generate automatically from product catalogs or enterprise terminology databases than natural-language instructions, and it is better suited to version control and batch-by-batch testing.
The biggest risk stems from the release mechanism: Microsoft says no customer action is required because the model is updated automatically on the server side. As a result, identical API calls may produce different transcripts, latency, and entity preferences even without a software deployment. Microsoft has shown only two examples and has not published per-language word error rates (WER), test sets, load conditions, or the distribution behind the claimed threefold speedup. The multilingual Fast model documentation lists `zh-CN`, but there is insufficient evidence that Traditional Chinese, Taiwanese accents, and mixed Mandarin-English speech received improvements of the same magnitude. Production teams should retain a fixed audio test set and separately measure WER, proper-name recall, real-time tail latency, and erroneous substitutions caused by phrase lists. They should also record each request's region and timestamp; otherwise, regressions will be difficult to reproduce after a server-side update.