Audio ingestion

Audio files use the same ingestion and refresh pipeline as documents. Import a recording with Add to memory → Audio on the dashboard, select its server-side path and choose Run. It runs in the existing queue with progress and Cancel. The file must be inside an approved workspace root, as for documents.

imprint ingest /path/to/meeting.m4a
imprint ingest /path/to/recordings
imprint refresh /path/to/recordings

Supported audio extensions are WAV, MP3, M4A, AAC, FLAC, OGG, Opus, AIFF/AIF and WMA. Folder ingestion includes them while audio.enabled is true (the default). This version accepts local file paths, not audio URLs or browser uploads. Video files and automatic speaker diarization are not implemented.

Models and languages

Settings → Audio ingestion controls the model, language and file limits. The default Large v3 downloads approximately 2.88 GiB once and is the most accurate: on OpenAI’s published figures it makes about a third fewer mistakes than Turbo in Romanian (Common Voice 15), and on a CPU it takes about as long as the audio, roughly twice Turbo’s time (ADR 0178). Large v3 Turbo (approximately 1.51 GiB) is about twice as fast and a little less accurate. The lighter Small model downloads approximately 464 MiB, with lower observed accuracy on the Romanian smoke fixture. Opening the dashboard/settings never starts a download; using an uncached audio file does. Weights are shared across workspaces and stored under models/ in the Imprint data directory. Recordings and decoded PCM are never copied into that model cache.

All three are multilingual Whisper models, run locally on CPU with int8 inference. PyAV’s wheels include the decoder libraries: no system FFmpeg, compiler, CUDA toolkit or microphone permission is required. Network access is used to fetch the selected pinned model on first use; transcription itself uses local files only. Missing Python runtime packages produce an instruction to run imprint bootstrap. Release installation/update installs the locked runtime dependencies; model weights remain separate downloads.

Auto detects the language of each decoding window. Romanian (ro) and English (en) pin it explicitly. Output preserves the recognized language; it is not translated into English. Auto can also recognize other languages supported by the multilingual model. Mixed-language content within one window can still be misrecognized; the reported language describes that window.

imprint config set audio.language ro
imprint config set audio.model large-v3-turbo
imprint config set audio.enabled false

Transcripts and source reading

Each recognized segment becomes a timestamped conversation event with a neutral speaker role, language and relative start/end seconds. Unidentified speaker means the recognizer has not identified who spoke. It never invents alternating user/assistant turns, speaker names or a recording date. The source reader shows the recording filename, segment text, language and time ranges. Original media playback/seeking is not supplied by this version.

Transcription is an estimate of the recording, not an exact transcript. Native checks recognized Romanian and English, but a Romanian fixture still contained a numeral error even with the larger model. Review names, numbers and important quotes against the recording before relying on them. These small functional checks do not establish multilingual accuracy or speaker separation.

Meeting recordings made on the dashboard’s Recordings page are transcribed by the same pipeline and, unlike imported audio files, tell voices apart within each channel and name them (ADR 0180), and leave out sentences Whisper likely invented (ADR 0181).

The pipeline redacts transcript text before indexing and stores the canonical transcript in the existing source snapshots. That completed snapshot is also the transcript cache, scoped to its workspace and the file hash/transcription profile. Identical files can reuse a completed transcript in the same workspace. Changing bytes, model, language or relevant runtime/settings invalidates reuse. imprint ingest ... --force bypasses the transcript cache and transcribes again. Deleting/purging a source removes that source’s snapshot, so it cannot supply a cached transcript afterward. Another retained source for identical bytes may still have its own snapshot. Raw media is not added to memories or sync; its original location remains provenance.

Limits, progress and failure

The defaults are 1,024 MiB per file and 120 minutes of decoded audio. Change audio.max_file_size_mb and audio.max_duration_minutes in Settings or with imprint config set. Limits must be positive. Decode duration is checked against metadata and actual decoded samples.

The decoder keeps 30-second windows in memory, with a one-second overlap. Word timestamps remove repeated words at window joins. This bounds decoded-waveform memory; transcript text is limited to 2,000,000 characters and 50,000 segments. One transcription model runs at a time in the ingest process. Progress shows actual model-download bytes, model loading, and transcription seconds; unknown duration uses an indeterminate indicator. Cancellation uses the queue’s existing owned process-group termination, including its force-stop escalation for native inference.

Failed decoding, model loading, checksum checks or transcription retain the prior indexed source and make ingestion fail visibly. No partially generated transcript is published. No recognized speech is an explicit “nothing to index” result; previous chunks are retained under the existing document no-text policy.

The decoder accepts audio demuxers and local file/pipe protocols only. Playlist demuxers and external MP4 data references are disabled. Model assets are revision- and SHA256-pinned; interrupted downloads remain partial model files and are replaced on retry. A corrupt completed weight file produces a repair error and is never loaded.

Validation

Run the ordinary Python tests, plus an optional native check with your own short English and Romanian fixtures:

PYTHONPATH=. /path/to/runtime/python scripts/smoke-audio-ingest.py \
  --english /path/to/english.wav --romanian /path/to/romanian.wav \
  --model-cache /path/to/dedicated/transcription-model-cache

The native check uses a disposable data directory, verifies model assets and checks both languages plus a mixed recording across a window boundary. It does not save or print transcripts. Run with each operating system’s native Python before claiming that platform validated. Linux/WSL execution and remaining gates are recorded in platform validation.

ADR 0095 records the source/caching semantics. ADR 0096 records the runtime and model policy.