Voiceprints & emotion

Enroll a voiceprint

1. Open Voiceprints, enter a display name and choose a matching speaker-embedding model. After changing models, older profiles may be incompatible and need re-extraction.
2. Add 1–8 independent reference recordings from one person: mono 16 kHz WAV, 1–60 seconds each. Inspect them after selection and remove damaged, constant or duplicate files.
3. Listen before creating a profile. Rename, archive, restore, review selected references and retain version history. New profiles are not used for automatic identity matching before calibration.

123

1. 1Enter a display name for enrollment.
2. 2Select the matching voiceprint model.
3. 3Add reference recordings from the same speaker.

Review speech activity

Optionally use Silero VAD or MarbleNet VAD to inspect speech intervals. Level measurements do not prove speech or a single speaker, and VAD hints do not automatically trim or reject recordings. Listen to a candidate span and explicitly adopt it for enrollment or comparison; you can undo and return to the original.

Compare and validate

Choose another independent recording as the probe and compare compatible profiles using the same model and runtime version. Inspect candidate scores, the top-two margin and excluded profiles. If results are ambiguous, uncalibrated or incompatible, follow the displayed status rather than treating the top candidate as confirmed identity.

Use independent material for verification and evaluation, not enrollment recordings as the test set. Review false accepts, false rejects, unknown speakers and actual microphone conditions. Reassess calibration and evaluation scope after changing models, references or conditions.

Read the sample source before the scores

Expand Voiceprint evaluation and choose Open report. Read the Sample source · report declaration card before comparing correct-identification or unknown-error rates. The report provider declares this source; opening the report does not verify who appears in the recordings.

• Synthetic fixture: checks functionality and cannot establish recognition quality for real people.
• Perturbation or conversion test: shows behavior under conditions such as noise or conversion. It cannot replace an independent test with human recordings.
• Original human recordings: the report declares that the recordings are human. Speaker identity, collection conditions and sample independence still need verification.

For Devices held out across dataset splits, Declared means that calibration and test sets use different device labels. It does not establish that the same person will be recognized after changing microphones, and it does not certify quality. Not held out means that these sets were not separated by device.

Older reports may lack source or device information. The app shows Not provided for each missing item; this does not establish human recordings or device separation. Ask the report provider for that information before using the scores to judge model quality. Opening a report does not change the current model or identity threshold.

Try the selected voice before dictation

These voice-trial steps require the Windows r512 or later installer preview. A trial does not establish real voiceprint quality or acceptance of automatic dictation delivery.

In Setup → Local models, register and select the embedding model and speaker segmentation model, then save. These are different from the optional speech-activity detector. The trial uses the saved configuration; no model path is entered here.

In the voiceprint library, select an unarchived profile. Under Compare a new recording, choose an independent 1–60 second recording, prepare the audio and complete its recording check. Do not reuse enrollment audio. Click Try the selected voice.

The trial checks the entire prepared recording, independently of any speech-activity crop selected above. It shares the actual dictation gate: single-speaker coverage, whole sentence, windows and speech intervals. It stops at the first rejection; untested intervals are not counted as passed.

Expand the interval details to see whether a rejection came from insufficient coverage, multiple speakers, low similarity or a small candidate margin. Low similarity does not prove it is another person. A passed trial is not identity authentication or a measured accuracy rate.

Cancel while processing to discard its result. Choosing another recording or profile, updating a profile or changing registered models invalidates the previous result. Model/profile mismatch: use Re-extract with current model in the profile; missing models: return to Setup.

The trial types no text and changes no threshold; automatic typing for a selected voice remains unavailable. A passed trial is not identity authentication and does not establish cross-application input acceptance. Windows r514 or later strengthens protection when changing target fields and stopping recording, but does not open this feature or lower voiceprint thresholds. See recording finalization and saved-audio checks for the steps after Stop. If your own independent recordings fail, improve the enrollment recordings and repeat the trial.

Verify source recordings

Expand Evaluation report in the voiceprint workbench and open a verification report. Opening a report recalculates statistics without reading recordings. Explicitly select Choose audio mapping and verify, then choose a JSON mapping.

1. Use schema_version: 1 and an audio_files array. Each entry contains sample (the exact id in input.manifest.samples) and path. Include every source sample; missing, duplicate or unknown IDs are rejected.
2. Use absolute paths or relative paths within the mapping directory. WAV files must be mono 16 kHz and 1–60 seconds. Verification does not convert, pad, upload or change the originals.
3. Review checked/matched/total counts and issues. Missing, damaged, hash-mismatched and duration-mismatched recordings identify the sample and file. Cancellation is supported; the last completed result remains explicitly labeled.

Scope of verification

If the report changed, reopen it before verification. Matching all recordings establishes only the bytes and format read at the recorded time. It does not validate speaker labels, model provenance, scores or recognition quality, and it does not install an identity threshold.

Meeting speakers and emotion

Review timeline speaker labels manually; voiceprint candidates can assist matching. Inspect segmentation labels separately from calibrated identity conclusions. Emotion detection depends on the selected analysis backend and available model and requires checking the audio. Emotion controls in voice synthesis do not automatically produce detection results.

First enrollment: inspect the recordings

Register a compatible speaker-embedding ONNX once in Setup → Local model library and save. In Voiceprints, enter a display name and select that entry. Enrollment, comparison and meeting review share the model choice; optional VAD uses its own saved choice. Manage models returns to Setup. Transcription, segmentation and emotion models do not substitute.

Add 1–8 independent recordings from one person, each 1–60 seconds. Choose WAV, MP3, FLAC, OGG or M4A, then explicitly Import & prepare each clip. Check the original format and derived mono 16 kHz PCM16 WAV before signal inspection. Sources stay unchanged, without trimming or padding. Remove damaged, constant or duplicate entries. A clip containing silence with an isolated click cannot support a voiceprint reference; choose continuous single-speaker speech. Review low-level warnings by listening. Create requires a model, name and valid prepared audio. For comparison, prepare an independent recording first; VAD is optional.

Cancel a processing job when needed and wait before changing inputs. After changing models, select a profile and re-extract with the current model before comparing independent audio. Earlier revisions remain; re-extraction is not calibration.

Independent acoustic emotion analysis

This tool needs an exported emotion2vec+ base/seed classifier ONNX with nine class labels and model metadata, such as classifier.onnx. The original model.pt, its model folder, a feature-only ONNX and a speaker model cannot substitute. If you only have source weights, independent emotion analysis is still waiting for a compatible export. Obtain that file from the model provider or administrator; preparing the export is separate from ordinary app use. Downloading source weights does not convert them automatically. With a compatible classifier prepared, the app analyzes locally through Rust and ONNX Runtime without Python; Python is only used by maintainers for the initial export.

In Setup → Models and parameters → Local model library, set Use to Emotion recognition and browse to the model file. For an original model.pt, use Check preparation to review missing files. The check does not save settings, load a model or download files. Register a compatible nine-label classifier.onnx and save Setup, then open Voice Tools → 06 Emotion and choose the registered model and a 0.25–10 second mono 16 kHz WAV. Source checkpoints, speaker models and feature-only ONNX files cannot replace the classifier. A file check does not establish successful inference.

1. In Setup guide, open the Local model library, choose Use → Emotion recognition and the emotion2vec family, then browse for the exported classifier.onnx file. This entry uses file selection. Give it a clear name, register it and save Setup.
2. Open Voice Tools → 06 Emotion and choose the saved classifier. If an older entry points to model.pt or a folder, the app retains it and explains why Analyze is unavailable. Use Manage models, edit that entry's path in Setup, save and return to the analysis tool. Do not rename model.pt to .onnx.
3. Choose 0.25–10 seconds of speech in WAV, MP3, FLAC, OGG or M4A. Explicitly click Import & prepare, check the original format and derived mono 16 kHz PCM16 WAV, then click Analyze sample. Sources stay unchanged, without trimming or padding. Start with two CPU threads. The .onnx suffix alone does not prove compatibility or successful inference; the app still checks the classifier's labels, metadata and output.

Use Cancel analysis to stop a running task; leaving the panel does not cancel it. Release model frees its memory after use. Review candidates against the recording and preserve uncertainty. Scores are uncalibrated, not emotional intensity or a conclusion about someone's internal state; analysis does not automatically change meeting labels or synthesis parameters. For cloned expression, follow Speech & voices.

1234

1. 1Choose a registered classifier; this example still selects original weights.
2. 2Use Manage models to edit the entry in Setup and choose the classifier.onnx file.
3. 3Original weights and model folders cannot be analyzed directly; their entries remain.
4. 4Analysis is blocked for an incompatible input; the original weights are not loaded.