When we introduced Post-Stream Refinement earlier this year, it closed the oldest gap in real-time speech: you could finally get instant streaming results and a highly accurate final transcript, with no latency penalty. But it kept one hard requirement — you had to tell the service, up front, which single language to expect. Real-world speech does not work that way. People code-switch mid-sentence, product and brand names cross languages, and a global app serves users who simply speak differently from one session to the next.
Today we remove that requirement. Multilingual Post-Stream Refinement enters public preview for Azure AI Speech in Microsoft Foundry, and for the first time ever a single real-time stream can transcribe multiple languages in one session — the spoken language is detected automatically, no locale is declared in advance, and the final transcript is refined for accuracy. Everything you already know about Post-Stream Refinement still applies; what changes is that the refinement pass itself is now multilingual. 📖 Read the Documentation
What’s New in This Release
If you have already used Post-Stream Refinement, here is exactly what changes with the multilingual preview — and what stays the same:
What’s new: the same real-time contract, now multilingual. One multilingual model runs the refinement pass, adding automatic multi-language detection, code-switching within a single utterance, open-range setup with no candidate list, and included diarization — while first-token streaming latency stays unchanged.
Quality Impact
In internal testing and partner evaluations across Tier-1 locales, multilingual Post-Stream Refinement reduced word error rate (WER) by approximately 10% relative on average, with double-digit relative reductions on the hardest cases — long utterances, proper nouns, and multilingual or code-switched speech. Partial-result latency is unchanged; only the final transcript is refined.
Gains are relative reductions versus the standard real-time model and vary by language, acoustic conditions, and content type. The refined final result may add a small amount of latency to the final segment; partial results are unaffected.
Measured quality gains (relative improvement vs. standard real-time transcription): about 10% relative reduction in average word error rate across Tier-1 locales; double-digit relative reduction on long utterances; double-digit relative reduction on proper nouns and named entities; no change to partial-result (first-token) latency. Relative reductions in word error rate; public preview, exact figures vary by locale and audio.
Supported Languages and Regions
The public preview supports 15 Tier-1 locales. Because language is detected automatically, a single stream can contain any mix of them:
Supported languages: 25 auto-detected languages spanning 29 market locales — any mix of which can appear in a single real-time stream without declaring a locale up front.
Available in these Azure regions:
Availability: multilingual Post-Stream Refinement is live in six Azure regions at public preview — East US, West US, North Europe, Central India, Southeast Asia, and Japan East.
Real-World Impact
Preview customers across industries — including travel, consumer electronics, automotive, aviation, and media — have reported positive gains in transcription quality. Customers testing multilingual and domain-specific audio have observed the clearest improvements on the hardest content: proper nouns, code-switching, and long-form speech. Several are actively validating the feature on their own audio ahead of general availability.
Get Started
Enabling multilingual Post-Stream Refinement is a small configuration change on your existing SpeechConfig. You will need:
- Speech SDK 1.50 or later. Earlier versions do not support the multilingual path.
- A Speech resource in one of the supported regions listed above.
- Auto-detect language configuration (open range) so the service identifies the language from the audio — no candidate list required.
Set the post-processing option to PostRefinement and pass an open-range AutoDetectSourceLanguageConfig when you create the recognizer. Here is a complete, copy-paste Python example, including the optional end-of-utterance detection line:
speech_config = speechsdk.SpeechConfig(
subscription=”YourSpeechKey”,
region=”YourSpeechRegion”)
# 1) Refine the final transcript (Post-Stream Refinement)
speech_config.set_property(
speechsdk.PropertyId.SpeechServiceResponse_PostProcessingOption,
“PostRefinement”)
# 2) Multilingual auto-detect – no candidate language list needed
auto_detect_config = speechsdk.languageconfig.AutoDetectSourceLanguageConfig()
audio_config = speechsdk.AudioConfig(use_default_microphone=True)
recognizer = speechsdk.SpeechRecognizer(
speech_config=speech_config,
auto_detect_source_language_config=auto_detect_config,
audio_config=audio_config)
💡 Tip: Refinement matters most for applications that store or process the final transcript — meeting notes, call analytics, compliance archives, AI summarization. If you only use partial results for a live display and discard them, your real-time UX (already fast) is unchanged, while any final transcript you keep improves.
Try Multilingual Post-Stream Refinement Today
Turn on higher-accuracy, language-aware transcription in your Azure AI Speech applications with a single configuration change. Available now in public preview in Microsoft Foundry.
We would love your feedback. Try Post-Stream Refinement in your applications and tell us how it improves your transcription quality.