LISTENER-BENCH

LISTENER-BENCH evaluates how real-time speech models listen during natural conversations. We stream 90-second recordings of real human speech to models and compare when and how much they speak with the original human listeners.

Blog post and full results: https://www.theliva.ai/blog/good-listener-bench

Two settings

The benchmark uses two datasets collected through our apps. Both contain real, unscripted English conversations, with each participant recorded on a separate audio channel. We select 90-second segments where one person does most of the talking and the other mostly listens.

config	source	windows here	sessions	what it is
peer_support	Emo-com	112	13	someone shares something difficult in their life; the other person listens
friends_chat	Yapdo	127	31	casual conversation between friends

This dataset release contains 239 segments. Segments containing sensitive content, including abuse, self-harm, graphic events, personal health or legal information, or mentions of people by name, have been excluded.

The full evaluation dataset used in the blog contains 241 peer-support and 184 friends-chat segments. The complete dataset, including excluded segments, may be available upon request for approved research collaborations.

Files

<config>/speaker_audio/<window>.wav   90-second speaker audio
<config>/reference.json               Speaker and human listener voice-activity segments and reference metrics
<config>/windows.json                 Segment IDs, pseudonymized session IDs, timestamps, and speech durations
<config>/transcripts/<window>.json    Word-level transcript of the speaker audio (ElevenLabs Scribe v2), window-relative timestamps
score.py                              Evaluating model outputs
requirements.txt                      Tested package versions

Session and segment IDs are pseudonymized.

How to run your model

For each segment, we stream the speaker's audio to the model in real time. The model does not hear the original human listener.

We record the model's responses and compare them with the human listener on the same speaker audio.

1. Stream each speaker recording to your model as live microphone input, using its default configuration. Document any additional prompts.
2. Record the model's audio output, aligned with the input.
3. Save one .wav per segment, matching the input duration.
4. Run the evaluation script: python score.py --config peer_support --outputs

Requirements: Python 3.10+ and the packages in requirements.txt (pip install -r requirements.txt). The silero-vad package ships its model weights, so no download happens at run time. The script refuses to score a partial set of windows unless you pass --allow-partial.

Metrics

We use voice activity detection to measure:

• Backchannels: Short listener responses under 1 second and at most 2 words, while the speaker keeps the floor (speaker speech continues through the response or resumes within 1 second of its end).
• Full turns: Listener responses that are not backchannels, measured per 90 seconds.
• Talk share: Percentage of the recording during which the listener speaks.
• Response latency: Time from the end of the most recent speaker stretch to the start of each full turn, reported as the median per segment.
• Mid-speech turns: Percentage of full turns that begin while the speaker is still talking (pooled over all full turns).

To apply the two-word backchannel definition, provide word-level transcripts of your model's outputs with --transcripts (the speaker transcripts in /transcripts/ use the same format). Each .json should contain word timestamps in the format {"words":[{"w":"yeah","s":1.0,"e":1.3}]}. Without transcripts, backchannels are identified using audio duration and speaker activity alone.

Voice activity is detected with Silero VAD (minimum speech 150 ms, minimum silence 250 ms, 60 ms padding); the same settings are used for the human reference and for model outputs.

Results

We evaluated GPT-Live 1, Gemini 3.8 Live, and PersonaPlex. The following results are from the 239 publicly released segments, using default model prompts. Values are medians over segments; the mid-speech column is the share of all full turns.

Peer support (112 windows):

	backchannels/min	full turns per 90 s	talk share	latency	full turns started mid-speech
Human listener	2.2	4	15%	1.9 s	28%
GPT-Live 1	5.2	4	15%	2.1 s	38%
Gemini 3.8 Live	1.1	4	12%	2.2 s	61%
PersonaPlex	2.7	6	25%	0.5 s	36%

Friends chat (127 windows):

	backchannels/min	full turns per 90 s	talk share	latency	full turns started mid-speech
Human listener	3.7	7	18%	1.5 s	25%
GPT-Live 1	4.0	8	22%	1.5 s	31%
Gemini 3.8 Live	1.6	6	17%	1.2 s	52%
PersonaPlex	2.4	9	27%	0.4 s	24%

Full-set numbers and the role-prompt conditions are in the post.

Access and terms

The dataset is gated and available for research use only. Redistribution and attempts to identify participants are prohibited. Data must be deleted upon request. All recordings are from real conversations collected with consent for research use. We also maintain a private held-out evaluation set containing sessions not included in this release. Researchers can submit model outputs or an endpoint for evaluation. For dataset access, evaluation, or research collaborations, contact contact@theliva.ai.