◫ LAION / AUDIO RESEARCH

CARBON / THREE THINKING SETTINGS / THREE SOUND FAMILIES

Every voice.
Every layer.

Follow the spoken words alongside the sounds and music around them. Compare new Carbon High, Low and Off annotations of the same six recordings, with utterance transcripts and up to three overlapping tracks in each category.

Speech + transcriptsSound effects & ambienceMusic18 new model responses

All boundaries and words are model predictions. Speaker characteristics are voice-based impressions, not verified demographics. No external speech recognizer or CLAP alignment is used.

ONE SHARED PLAYER / SYNCHRONIZED COMPARISON

Listen across the layers.

Loading experiment…

RECORDING

Loading…

0:00 / 0:00
Press Play to load the recording.
Click any event to hear its original interval.

THINKING COMPARISON

VISIBLE LAYERS

Speaker colors stay consistent within a response. Speaker IDs are not matched between settings.

Each category has up to three visual tracks. A track can carry different speakers at different times; the speaker badge and color identify the voice. If more than three events overlap, every extra prediction remains in a labeled list below its category. Nothing is silently removed. These are annotations over the original recording, not separated audio stems.

READ THE EXPERIMENT

Words, voices, and the scene around them.

01

One complete recording

Carbon receives the complete audio and the same new prompt for all three settings. It predicts events and timestamps directly, without candidate captions, reference transcripts or subsequent alignment.

02

Three independent categories

Speech contains natural utterances, original-language transcripts and anonymous speaker IDs. Sounds contains non-musical effects, ambience and non-speech vocalizations. Music follows audible musical passages, including music under speech.

03

Compare what you hear

Play a recording or individual interval. The current utterances and concurrent sounds update together. Check missing words, false sounds, speaker changes and boundary placement. More events or more tokens do not establish better accuracy.

Transcripts, speaking style, emotion and speaker attributes

Each speech event carries the original-language transcript, language, speaking style, expressed emotion, broad perceived age and perceived gender. The caption explicitly describes these voice impressions. They are unverified acoustic judgments, not identity, actual age, gender identity or a person's true feelings. “Uncertain” is a valid prediction.

Partial and unintelligible utterances are labeled; unclear words should appear as [inaudible]. “Verbatim” is the requested transcription format, not a guarantee that the model heard every word correctly. Singing belongs in music; laughs and other non-speech vocalizations belong in sounds. No external transcription check has been applied.

How the overlapping tracks are arranged

Speech, sounds and music each receive at most three concurrent lanes. Original events are processed by onset and placed in a free lane, preferring the previous lane for that speaker when it is available. A lane is a display position, not a permanent speaker identity. Separate categories may overlap freely.

When all three lanes in a category are occupied, the additional event stays in an explicit overflow list, with its unchanged transcript, caption, timestamps and playback button. This is a presentation limit, not an inference filter. Original out-of-range timestamps are flagged; only their drawn bar is clipped to the recording. Invalid intervals are retained but not played.

High, Low and “None / Off”: the exact API settings

High and Low use the Interactions API with their corresponding thinking level. “None / Off” uses generateContent with an explicit thinking budget of zero: Interactions rejected the literal none enum in the previous experiment. All three use temperature 1.0, a 32,768-token output limit, identical full WAV bytes, the same prompt and the same JSON schema.

Off changes the API method as well as the thinking setting. There is one sampled response per clip and condition, with no fixed seed or human ground truth. The earlier foreground/background results remain available under Earlier comparisons; they used a different prompt and should not be treated as a controlled comparison of this new task.

Exact prompt and output schema

Only the measured duration of the clip is appended to the shared prompt. Credentials and audio payloads are never published.

Measured usage, elapsed time and estimated cost

Client elapsed time and token usage are measured for these new annotations. Dollar amounts are calculated from usage and the published standard API rates through November 30, 2026, including reported retry usage; they are not invoice charges. Batch-rate projections and the December rate period are included in the downloadable manifest. No Batch job was submitted. Pricing source ↗

Manifest, prompt & measurements ↓ · Per-clip usage · CSV ↓ · Reproduction README ↗