How to Use pyannote Speaker Diarization Precision-2

Before you start

pyannote speaker-diarization-precision-2 is a speaker diarization pipeline distributed through the pyannote Audio library. You normally provide an audio file and receive time segments labeled by speaker, such as SPEAKER_00 or SPEAKER_01. The labels are local to each file and are not linked to named identities unless you provide your own enrollment or mapping step.

The process runs locally through Python when you use the pyannote Audio package, but pipeline weights are normally downloaded from a model repository. Access to the pipeline model may require accepting the model license and supplying a token. Check the official pyannote Audio model page or your installed package documentation for the current access requirements instead of assuming a local-only or fully offline workflow.

Prepare an audio file in a supported format, ideally with clear speech and limited background noise. The input device used for recording does not need to match the device used during model development, but poor input level or heavy room echo can still affect segmentation quality. A higher sample rate is not a guarantee of better diarization; most pipelines resample internally to the expected rate.

Step-by-step instructions

Install the required Python packages in a clean virtual environment so you do not change system-wide dependencies. The exact package versions may change, so use the installation command shown in the current pyannote Audio documentation for your operating system.

python -m venv diarization-env
source diarization-env/bin/activate
pip install pyannote.audio

Obtain a model access token through the authentication method required by your pyannote Audio version. Some versions accept a token directly in the pipeline call, while others require an environment variable or configuration file. Do not paste a token into shared code or logs.

from pyannote.audio import Pipeline

pipeline = Pipeline.from_pretrained(
    "pyannote/speaker-diarization-precision-2",
    use_auth_token="YOUR_TOKEN_HERE"
)

Run the pipeline on your audio file. Use a variable or environment setting for the token in real scripts instead of leaving it in the source code.

diarization = pipeline("meeting.wav")

for turn, _, speaker in diarization.itertracks(yield_label=True):
    print(f"{turn.start:7.2f} {turn.end:7.2f} {speaker}")

For a long recording, start with a short test clip that contains two or three speakers and known speech boundaries. Work locally with a copy of the file if you need to change the format or trim it first. Preserve the original recording until you have confirmed the output.

If you use the pipeline in a call or live application, treat diarization as a processing step that follows the audio capture path. The output device used for playback is separate from the diarization result and does not affect the speaker segments.

Check the output

Inspect the printed turns to confirm that the pipeline returned time ranges and speaker labels instead of an empty annotation. Compare a few segments against the original audio by listening at those timestamps and checking whether speech belongs to the same speaker.

Watch for known limitations instead of expecting exact word-level boundaries. Short interjections, laughter, overlapping speech, and speakers recorded at low input level may be missed, merged, or assigned to a new label. Background noise can also split one speaker into several labels.

diarization = pipeline("meeting.wav")

for turn, _, speaker in diarization.itertracks(yield_label=True):
    duration = turn.end - turn.start
    print(f"{speaker}  {turn.start:.2f}s -> {turn.end:.2f}s  ({duration:.2f}s)")

If the file is stereo, confirm that the pipeline accepted it or resampled it as expected. Some pipelines require mono input or automatically convert the input. Check the current model documentation if the script fails on a stereo recording.

Common problems and next steps

A common error is an authentication failure when the model cannot be downloaded. Verify that you have accepted the model license and that the token has access to the correct repository name. The model ID is often written as pyannote/speaker-diarization-precision-2, so check for a missing slash or extra space in the identifier.

If the result contains too many speaker labels, try reducing frequent speaker changes by cleaning the audio or using a short segment with one speaker per turn. Some pipeline versions expose a minimum duration or clustering parameter, but do not assume a setting exists unless it is documented for the installed version.

If the result merges speakers, check whether the recording has very similar voices, low input level, or one microphone far from the speakers. Improving recording position usually helps more than changing the sample rate after capture.

Keep the original audio and the segment output together. For privacy-sensitive recordings, process files on a machine you control and remove or secure temporary copies after you have verified the result.