Audio Annotation Tools Compared: ELAN, Praat, Label Studio and Potato

Audio annotation attaches labels to stretches of a recording: who is speaking, what they said, which sound occurred, or how good the recording is. ELAN and Praat are the standard desktop tools for one linguist working through tiers of annotation. Potato, Label Studio and Prodigy run in a browser, which suits studies where many annotators label the same recordings. Potato also runs the study around the labels: it recruits listeners on Prolific or MTurk, trains and screens them, and reports agreement on segments, with no paid plan. The browser tools also differ in whether they read existing transcripts.

Speaker diarization splits a recording by who is speaking. Speech recognition produces a transcript, which annotators then correct. Sound event detection marks where a sound such as a siren or a door occurs. A tier, in ELAN and Praat, is one layer of annotations over the recording's timeline, so words, phonemes and gestures can be labelled separately. Mean opinion score rates the perceived quality of a clip.

This page covers audio and speech. Annotation Tools Compared covers every data type on one page.

Which audio annotation tools are worth comparing?

Which tool for which audio task?

Running a listening or speech study

The table above is about the labelling. When many people label the same recordings, the tools differ more:

A dash means we found nothing for it in the tool's documentation in September 2026. If a cell is wrong, tell us.

Where Potato is ahead for audio

Listening tests with crowdworkers. Quality ratings and perception studies are often run with crowdworkers rather than a lab team. Potato brings listeners in from Prolific or MTurk, gives them practice clips with feedback before the real ones, mixes in gold clips and attention checks, and shows a completion code at the end. See Running a Study on Prolific and MTurk and training phases.

Agreement on segments. ELAN compares two tiers at a time. Potato compares every pair of annotators on how much their segments overlap and whether they found the same segments, at one matching threshold. Segment agreement does not yet have a chance baseline, and its documentation says so.

From ASR output to ELAN. Annotators start from an existing transcript in one of 21 formats, or from Potato's local transcription and diarization, and the result exports to EAF and TextGrid for whoever continues in ELAN or Praat.

Starting from an existing transcript

Most speech projects start with a transcript from Whisper, a cloud API or a subtitle file. Potato reads 21 transcript and subtitle formats and renders their turns as speaker bubbles synced to the audio, so annotators correct segments instead of typing them from scratch. The formats include Whisper and WhisperX JSON, AWS Transcribe, Deepgram, AssemblyAI, Rev.ai, SubRip, WebVTT, NIST CTM, Praat TextGrid and ELAN EAF. Potato does not parse RTTM diarization files on their own, Transcriber, EXMARaLDA or CHAT, and those need converting first. See Transcript Formats.

Since version 2.9 Potato can also transcribe and diarize on your own machine, using faster-whisper and sherpa-onnx, with no PyTorch and no Hugging Face token. See Transcribing locally.

In Label Studio and Prodigy, a transcript from another system is converted into the tool's own task format first. Label Studio's Audio tag pairs with a TextArea for transcription, and Prodigy's audio.transcribe recipe combines playback with a text box.

Moving tiered annotation between ELAN, Praat and Potato

Potato's tiered_annotation scheme holds independent and dependent tiers, so a word tier can subdivide an utterance tier in time, as in ELAN. Potato exports EAF and TextGrid. A study can collect annotations from many people in a browser and hand the result to someone who works in ELAN or Praat.

ELAN remains the better tool for one expert working through a recording in detail, with synchronized views of up to four videos. Praat is for acoustic analysis, which Potato does not attempt.

Agreement over time

Two annotators marking the same sound rarely agree to the millisecond, so agreement on audio has to decide how close two boundaries must be to count as the same event.

• ELAN has a built-in inter-annotator reliability calculation. It offers a modified Cohen's κ, after Holle and Rein, which links annotations that overlap by a set minimum, and the Staccato algorithm, which accounts for chance by comparing a segmentation with randomly generated ones. It compares pairs of tiers.
• Potato compares every pair of annotators on temporal IoU, detection F1 and an overall agreement score that gives half credit to matched segments with different labels, all at one matching threshold. None of these has a chance baseline yet, and its documentation says so. See Agreement over Time.
• Label Studio keeps agreement for its paid plans.

An audio segmentation task in Potato

annotation_schemes:
  - annotation_type: audio_annotation
    name: sound_events
    description: "Mark each sound and choose its type."
    mode: label
    labels:
      - {name: speech, color: "#4ECDC4"}
      - {name: music, color: "#FF6B6B"}
      - {name: noise, color: "#F39C12"}
    zoom_enabled: true

Annotators drag across the waveform to create a segment and pick its label. Audio Annotation covers classification, transcription, MOS ratings and diarization in Potato.

What Potato does not do for audio

• No acoustic analysis. Pitch, formants and spectral measurements belong in Praat.
• Simpler alignment tools than ELAN's. ELAN's desktop interface gives finer control over time alignment.
• No parser for standalone RTTM, Transcriber, EXMARaLDA or CHAT files.

Checked against each project's documentation in September 2026.

Frequently asked questions

What is the best free audio annotation tool?

ELAN and Praat are free and are the standards for linguistic and phonetic work by a single annotator. For a study in which several annotators label the same recordings in a browser, Potato and Label Studio's Community Edition are free. Potato includes agreement metrics, which Label Studio keeps for its paid plans.

Does ELAN calculate inter-annotator agreement?

Yes. ELAN's inter-annotator reliability calculation offers a modified Cohen's kappa and the Staccato algorithm, and it compares annotations on pairs of tiers across files.

Can I annotate speaker diarization in a browser?

Yes. Potato reads diarized output from WhisperX, AWS Transcribe, Deepgram, AssemblyAI and Rev.ai, shows turns without a speaker as Unassigned with a picker, and from version 2.9 can diarize locally. Label Studio can also label speaker regions on a waveform.

Which annotation tools read Praat TextGrid and ELAN EAF files?

Praat and ELAN work in their own formats. Potato reads and exports both, so a project can move between a browser study and either desktop tool.

Do I need to transcribe audio before annotating it?

Only if the annotation is about what was said. Sound events, speaker turns and quality ratings can be labelled on the waveform directly. For speech content, start from an existing transcript when you have one, and transcribe from scratch when the transcript is the deliverable or the audio is bad enough that ASR output would mislead annotators.

Which audio annotation tools work with Prolific or Mechanical Turk?

Potato has both built in. A listener arrives from Prolific or MTurk with their participant ID in the URL, works through any consent, screening and training phases, and gets a completion code at the end. ELAN and Praat are desktop tools, and we found no crowdsourcing integration documented for Label Studio or Prodigy.

Further reading

• Annotation Tools Compared, for every data type on one page
• Text Annotation Tools Compared
• Image Annotation Tools Compared
• Video Annotation Tools Compared
• AI Agent Evaluation Tools Compared
• World Model and Robot Episode Evaluation Tools Compared
• Annotating Whisper Transcripts