01
Prepare an interview timeline.
Group audio intervals by speaker label so an editor can navigate the recording. Let the editor name speakers manually after review; a diarization label is not a verified identity.
pyannote Community-1
Detect who spoke when.
pyannote Community-1 detects who spoke when by assigning speaker labels to audio intervals. Deploy its diarization pipeline on AI Fabric to build interview timelines or add speaker labels to a separate transcript.
€5 in free credits for every new customer.
Plan your deploymentAI Fabric access is opening gradually. GPU configuration is confirmed during onboarding.
Speaker timelines for recorded conversations.
Help editors navigate an interview or attach speaker labels to transcript passages. Preserve the original timing and keep uncertain assignments reviewable.
02
Structure a podcast edit.
Use speaker turns to find exchanges and hand-offs in a long recording. Preserve overlaps and listen to uncertain boundaries before cutting audio at a predicted timestamp.
03
Enrich a searchable transcript.
Combine the diarization output with independently generated transcript segments. Keep the recording reference and timing on every passage so search results can jump back to the source.
Model capabilities and upstream setup: pyannote Community-1 publisher reference. Reviewed .
Deployment preparation
Run pyannote speaker diarization on a GPU.
Community-1 runs locally after its weights are loaded; pyannote.audio defaults to CPU unless you move the pipeline to CUDA. Arrange checkpoint access and pass your download token through the environment.
- Obtain access to the gated model and install a compatible pyannote.audio environment with its audio dependencies.
- Load the pipeline once and test a recording with known speaker turns, including interruptions and overlap.
- Serialize speaker intervals, then align them with transcript timing in a separate application step.
import os
import torch
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained(
"pyannote/speaker-diarization-community-1",
token=os.environ["HF_TOKEN"],
)
pipeline.to(torch.device("cuda"))
output = pipeline("interview.wav")
for turn, speaker in output.speaker_diarization:
print({
"start": turn.start,
"end": turn.end,
"speaker": speaker,
})Requires authorized model access, HF_TOKEN in the server environment and interview.wav. Verify decoding dependencies and keep the token out of client-side code.
GPU memory & performanceCapacity depends on recording duration, segmentation and the full pipeline. Measure peak memory and processing time on RTX A4500 or RTX 2000 Ada before setting upload limits or parallel job counts.
In your applicationKeep speaker labels scoped to the recording. A label such as SPEAKER_00 should not be treated as the same person across unrelated files.
AI Fabric availability & costRTX A4500 (20 GB) and RTX 2000 Ada (16 GB) are the GPU options. Access, runtime settings and billing are confirmed during onboarding. No Blade latency or throughput benchmark is published for this configuration.
Align speaker intervals with a transcript.
Transcription and diarization use different boundaries. Match intervals by overlap, retain ambiguous assignments and let a reviewer correct them. If the product requires a single speaker per instant, evaluate the pipeline’s exclusive diarization output for that alignment step.
01 / InputAudio recording
02 / ProcessSpeaker intervals
03 / ResultAlign transcript → review labels
pyannote Community-1
deployment FAQ.
Does Community-1 transcribe what people say?
No. It detects speaker intervals. Use a speech-to-text model such as Whisper for the words, then align the two outputs.
Can I provide the expected number of speakers?
The pipeline accepts an exact speaker count or lower and upper bounds. Supply them when known and test whether the assumption holds on real recordings.
Does a speaker label identify a real person?
No. Labels group speech within a recording. Your product needs a separate, reviewed process to attach names or identities.
How much does running pyannote Community-1 on AI Fabric cost?
Measure processing time per audio hour with realistic speaker counts and recording quality. Add transcription and alignment separately if the final product is a speaker-labelled transcript.
The GPU billing unit and idle policy are confirmed during onboarding. Managed Models API rates do not apply to this deployment.
Checkpoint: pyannote/speaker-diarization-community-1. Access requirements and licence: pyannote Community-1 publisher reference.