speech_core 0.2.0

Provider-neutral speech recognition, synthesis, and analysis contracts.

Changelog

0.2.0

• Add SpeakerEmbedding, a speaker vector carrying the (providerId, modelId, dimension) space that produced it, plus speakerSimilarity (null across spaces), requireSpeakerSimilarity (throws SpeechEmbeddingSpaceMismatch) and l2NormalizeVector. Comparing embeddings from two different models can no longer return a plausible number, which is what turns a silent mislabel into a detectable one. This is the provider-output contract and where this package stops: saved voiceprints, centroid blending, and nearest-profile search are cross-meeting identity policy and belong to the intelligence layer.
• Add SpeakerSegment.embedding, an optional field. Providers that only forward speaker labels leave it null; existing call sites are unaffected.
• Add SpeechCapability.speakerEmbedding and SpeechCapability.inverseTextNormalization, appended at the end of the enum. Persist capability sets by name, never index.
• Add the InverseTextNormalizer and CustomizableInverseTextNormalizer provider contracts plus InverseTextNormalizationRule, so spoken-to-written conversion has a provider-neutral seat instead of being reached through a direct adapter import.
• Add BatchRecognitionSegment.words, an optional List<SpeechWord> that defaults to empty and is copied on construction. Batch recognition could previously only return a block of untimed text, so nothing downstream could cut it into sentences, align it to a diarization span, or seek back into the audio — an asymmetry with the streaming path, which has always carried SpeechTranscript.words. Existing call sites are unaffected.
• Add SpeechAudioGuards with minimumRecognitionDuration (one second, the 16,000 samples at 16 kHz that batch recognizers need) and the speech_audio_too_short failure code. Providers refuse shorter audio with a typed SpeechFailure instead of returning an empty transcript, which is indistinguishable from the user having said nothing.
• Add the turn-completion seam: TurnCompletionScorer with scoreTurnCompletion(TurnCompletionRequest), plus TurnCompletionScore (probability, isComplete, optional threshold) and TurnCompletionAudio (16 kHz, an eight-second maximum window, a 0.5 default threshold compared strictly greater-than). The request carries a finite mono BufferedAudioSource — narrower than the AudioSource the other batch requests accept, because a scorer that cannot see the end of the window has nothing to score — plus an optional threshold override, cancellation, and provider options. This is what lets a turn detector reach a prosody model in an adapter package without the two packages depending on each other.
• Add SpeechCapability.turnCompletion, appended at the end of the enum. Breaking for exhaustive switches: code that switches over SpeechCapability without a default arm stops compiling until the new member is handled. Persist capability sets by name, never index.
• Add TranscriptInverseTextNormalizer, an optional refinement of InverseTextNormalizer adding normalizeTranscript(SpeechTranscript), plus the normalizeTranscriptPreservingWords helper implementing the baseline behavior over any string normalizer. ITN and word timings could not coexist: normalizing a transcript meant reducing it to a string and throwing away the timings that sentence segmentation, diarization alignment, and seeking all depend on. Deliberately a sub-interface rather than a new member on InverseTextNormalizer, so existing implementers keep compiling and a provider that can only rewrite strings stays a valid one; callers test for the stronger contract with is.

0.1.0

• Initial public release.