pub.dev

azure_speech_recognition_services

• Last release 4 days ago

  10 Nov 2026

• Ships on a steady schedule

  a new release about every 4 months

• Nearly every release is documented

  notes for 4 of 4 stable releases

• Nothing withdrawn

  no release was ever pulled

• 8 months old

  4 releases · first in 2026

1. 1.0.010 Nov 2026Deprecated

   A rewrite: a new stream-based API, new Android and iOS code, and web support. The 0.0.x API still works but is deprecated; see MIGRATION.md.

   A rewrite: a new stream-based API, new Android and iOS code, and web support. The 0.0.x API still works but is deprecated; see MIGRATION.md.

   New

   • AzureSpeech:
     • recognizeOnce(), startContinuous(), stop() (graceful), cancel(), prepare() (pre-connect) and dispose().
     • Streams: events (a sealed SpeechEvent), results, finalResults, audioLevel and errors.
     • state as a ValueListenable.
   • AzureCredentials: subscription key, tokens from a tokenProvider (refreshed inside the live recognizer), custom endpoint, custom host, and Entra ID tokens.
   • RecognitionConfig:
     • Language auto-detect (at-start and continuous) and phrase lists with weight.
     • Profanity, output format, word timestamps and dictation.
     • Segmentation silence timeout, semantic segmentation, and initial and end silence timeouts.
     • Stable-partial threshold and Custom Speech endpoint ID.
     • Speaker diarization, pronunciation assessment (with prosody and IPA), and stop phrases.
     • It is validated before anything is sent to Azure.
   • SpeechResult: text, offset and duration, detected language, speaker ID, confidence, lexical, ITN and masked-ITN forms, words with timing, pronunciation scores per text, word and phoneme, and the no-match reason.
   • Typed errors: AzureSpeechException with a SpeechErrorCode (mapped from Azure's CancellationErrorCode), the CancellationReason, and isTransient.
   • Built-in microphone permission (checkPermission, requestPermission, and an automatic request on start).
   • A mic level meter. The plugin captures 16 kHz mono PCM itself.
   • Web support through the official Azure Speech JS SDK 1.53.0, loaded on demand with Subresource Integrity, or self-hosted.
   • The platform interface (AzureSpeechPlatform) is public for fakes in tests.
   • Android: Speech SDK 1.53.0, R8 keep rules, and the Flutter 3.47 plugin Gradle template. Builds with both AGP 8 and AGP 9.
   • iOS: Speech SDK 1.52.0, the privacy manifest is bundled, and the podspec metadata is filled in.

   Fixed (bugs in 0.0.3)

   Dart:

   • A native event without a registered handler crashed: every handler was called with !.
   • The default language en-EN isn't an Azure locale. It is now en-US.
   • initialize(timeout:) threw a FormatException or a bare String. It now throws an AzureSpeechException.
   • initialize reset handlers on a throwaway instance. The channel handler was only installed once the singleton was constructed, so apps that only called the static methods got no events.
   • The static methods fired and forgot invokeMethod, so platform errors became unhandled async errors.
   • The segmentation timeout was ignored in continuous mode.
   • The pubspec declared Linux, macOS, Windows and web implementations that did nothing, and they threw MissingPluginException.

   Android:

   • stopContinuousStream never replied when nothing was running, so await stopContinuousRecognition() hung forever.
   • A double tap started two recognizers. The running flag was only set after start completed, and the first recognizer and its mic leaked.
   • lateinit recognizer access crashed when the flags got out of sync.
   • AudioConfig, SpeechConfig and recognizers weren't closed on every path. The mic stayed open, and nothing was stopped on engine detach or hot restart.
   • Exceptions from SDK futures were swallowed in a thread pool. The canceled event wasn't handled, so a wrong key or a lost network produced silence: an empty final result and no error.
   • The partial-result listener was attached after recognizeOnceAsync had started, so the first partials could be lost. "Recognition started" fired before the session had started.
   • The plugin didn't check permissions, so starting without RECORD_AUDIO failed obscurely.
   • Transcripts and pronunciation JSON were written to logcat.
   • The manifest had a deprecated package attribute, and MicrophoneStream.kt was dead code in the wrong package.

   iOS:

   • try! and force-unwrapping crashed the app on an empty or invalid key, region or config.
   • Platform-channel calls were made from background threads, which Flutter requires to be on the main thread.
   • startContinuousRecognition and stopContinuousRecognition blocked the main thread.
   • The audio session was set to .record and never deactivated. That silenced the app's own playback and other apps' audio.
   • Cancelling a single-shot recognition only set a flag; the mic kept listening until silence.
   • The task dictionary was mutated from several threads, a data race.
   • The canceled-result path force-unwrapped the cancellation details.
   • Transcripts were printed to the console.
   • The podspec had placeholder homepage and author, and the privacy manifest wasn't bundled.

Your coding agent can read these notes before it upgrades. Set up the MCP server →