Recursive AI
onespeechDocs

Changelog

OneSpeech 1.0.0CHANGELOG.md at 63002de Markdown

All onespeech docs

All notable changes are documented here, following Keep a Changelog.

[1.0.0] - 2026-11-10

First release. Measured results and limits: docs/results.md.

Highlights

  • Transcription with four model families behind one seam: Whisper (tiny to large-v3 and large-v3-turbo, the default), Qwen3-ASR 0.6B and 1.7B, FastConformer (Parakeet-TDT-0.6B-v3 and Nemotron 3.5 ASR streaming, by name), and Moonshine (the live default). VAD-first long-file chunking in constant memory, language identification, optional mixed-int8 variants.
  • Word timing from the best available method in a fixed order (native, Qwen3-ForcedAligner, omniASR-CTC, DTW), recorded per word; public Speech.Align for any text and audio.
  • Diarization: a tier-1 default of Silero VAD, CAM++ embeddings and spectral/AHC clustering with resegmentation, merged into the transcript. The default does not detect overlapping speech. Tier-2 upgrades, loaded by name from the publisher: Nemotron 3 Diarization (overlap-aware, up to eight speakers) and the gated pyannote segmentation-3.0 slot.
  • Speaker recognition: embeddings (CAM++, ECAPA-TDNN), enrolment, identification and verification with calibrated thresholds; in-memory only, never persisted.
  • Live sessions with one event model (Partial, Final, Endpoint, SpeakerTurn) on one audio clock; true streaming partials on Moonshine and Nemotron 3.5 ASR, utterance mode for the other families.
  • Contextual biasing through WithPhrases where the model supports it; other models refuse with a typed error.
  • Packages: RecursiveAI.OneSpeech, .Media (containers, SRT/VTT/ASS/TTML, caption muxing), .Cli (the onespeech tool), and offline weight packages .Models.CamPlusEnglish, .Models.MoonshineTiny, .Models.MoonshineBase, .Models.MoonshineStreamingTiny. NativeAOT and trim clean; no native code.
  • Models: tier-1 (MIT, Apache-2.0) models download on first use; tier-2 (CC-BY-4.0, OpenMDW-1.1, gated) models load only by name from the publisher's files and are never redistributed.

Changes since the last preview follow.

Added

  • The gated pyannote segmentation-3.0 model in the segmentation slot, with an overlap-aware diarizer over it (PyannoteDiarizer), loaded from publisher files the user is entitled to; never a default.
  • Mixed-int8 variants of Whisper large-v3-turbo, Whisper small and Moonshine streaming-tiny, selected by model ID (-int8) and converted locally from the same pinned weights; defaults remain fp32.
  • Contextual phrase biasing through SpeechRecognizer.WithPhrases: Whisper prompt conditioning, Qwen3-ASR context, and shallow fusion over a phrase trie for Parakeet and custom CTC/RNN-T/TDT models. Families that cannot bias refuse with CapabilityNotSupportedException.
  • Speech.Enroll, Speech.Identify, Speech.Verify and Speech.ClearSpeakers, with enrolled names on transcript speakers.
  • Speech.PrefetchAsync(language, model) fetches every model a call would need.
  • Path-input *Async overloads for every verb, and stream input for TranscribeStream and DetectLanguageAsync.
  • Uniform Load(ResolvedModel, ...) and LoadAsync(store, model, ...) on every downloadable family.
  • ModelUnavailableException, and ModelId on model errors.
  • TranscriptJsonConverter, so plain System.Text.Json writes the documented schema-1 shape; value equality for Transcript, Segment, Turn and ModelInfo.

Changed

  • Requires RecursiveAI.OneModel 1.2.0.
  • Faster CAM++ feature extraction with identical output.

Fixed

  • Transcript.ToJson and FromJson handle recordings up to 24 hours (documented limits, checked before allocating).
  • One language-tag rule for every entry point; unsupported languages are refused, and Align no longer returns untimed words when no aligner applies.
  • The role base classes implement IDisposable. Wrappers and sessions borrow, and never dispose, the models they use.
  • CreateLiveSession accepts any VoiceActivityDetector.
  • Consistent optional language and cancellationToken parameters, with the token last everywhere; MediaInput.TranscribeStreamAsync is renamed TranscribeStream.
  • WhisperOptions.Progress is IProgress<TimeSpan>; MoonshineOptions.IntraOpThreads is renamed Threads; thread counts default to 1.
  • Invalid in-memory PCM throws ArgumentException.
  • The Silero VAD provenance ID is the catalog ID silero/vad-v6.2.3.

[0.4.0-preview.1]

Added

  • Qwen3-ASR 0.6B and 1.7B recognition, language identification and context biasing, and Qwen3-ForcedAligner word timing.
  • CTC forced alignment with omniASR-CTC-300M and public Speech.Align for any text and audio.
  • Moonshine recognition with true streaming partials, incremental decoding with cached attention, and live sessions that keep up in real time on one core with no steady-state allocation in OneSpeech's own code per audio chunk.
  • Online speaker clustering for live sessions.
  • Parakeet-TDT-0.6B-v3 and Nemotron 3 Diarization by name, loaded from the publisher's files and converted locally. They are never defaults and are never redistributed. Nemotron 3 Diarization detects overlapping speech.
  • Gated publisher models: a user token is sent only to the publisher's own file URLs, with typed errors for missing tokens, rejected tokens and gates not yet accepted.
  • Custom CTC and RNN-T/TDT models of a known shape, decoded by the public Decoding primitives with native word timing.
  • NeMo mel features.
  • RecursiveAI.OneSpeech.Media: audio from any container OneMedia opens, caption export (ToSrt, ToVtt, ToAss, ToTtml) and caption muxing.
  • Whisper large-v3.

Changed

  • Requires RecursiveAI.OneModel 1.2.0-preview.2.
  • Qwen3 BF16 graphs are converted only when requested.

[0.2.0-preview.1]

Added

  • Live sessions with one event model (PartialEvent, FinalEvent, EndpointEvent, SpeakerTurnEvent): finals are immutable, partials carry a revision, and all events share one audio clock. Speech.Listen and Speech.TranscribeStream; utterance mode (voice-activity endpointed finals) for offline recognizers.
  • Chinese speech uses the Chinese CAM++ speaker embedder for diarization; other languages use the English one.
  • A diarization speech-activity profile fitted on development meetings.

Fixed

  • Transcript JSON writes times exactly at 100-ns resolution instead of rounding through binary floating point.

[0.1.0-preview.1]

Added

  • Shared speech contracts, the Transcript document model and its versioned, bounded JSON form.
  • Speech.Transcribe, TranscribeAsync and DetectLanguage: transcription of WAV, MP3, FLAC, Ogg, M4A/3GP/ADTS and AMR files, returning words with timings, confidence and timing method, segments, language and speaker turns.
  • Whisper recognition (tiny, base, small, large-v3-turbo) with greedy and beam decoding, temperature fallback, VAD-first long-form chunking, cross-attention word timing and language identification.
  • Embedded Silero VAD v6.2.3.
  • Speaker embeddings (CAM++ English and Chinese, ECAPA-TDNN), an in-memory speaker registry, and clustering diarization (spectral and agglomerative) merged into transcripts, including channel diarization for separated-party stereo. The clustering diarizer does not detect overlapping speech.
  • A pinned model catalog with verified download, local conversion on first use, a per-user cache, the ModelSource hook and onespeech models pull.

Known limitations

  • Whisper large-v3 was refused with a typed error in this release (fixed in 0.4.0-preview.1).