conversation-conditioned two-speaker pilot

Labnote 018: conversation-conditioned two-speaker pilot🔗

Question🔗

Can conversation-aware turn direction produce a convincing two-speaker exchange where isolated line direction does not?

Frozen pilot🔗

• One authored, low-stakes six-turn exchange between two alternating housemates.
• Two distinct Qwen3-TTS 1.7B CustomVoice preset speakers.
• An isolated arm giving every line the same generic conversational instruction.
• A conversation-aware arm giving each line an authored reaction and delivery direction

grounded in the shared scene.

• Identical text, speakers, per-turn seeds, generation settings, and silence gaps across

both arms.

• No candidate selection or speech editing; only deterministic silence was inserted when

assembling the generated turns.

• Blind review of the two complete conversations before arm identity was revealed.

The listener judged four whole-scene gates: natural individual voices, apparent reaction between speakers, conversational turn timing, and whether the complete exchange was convincing.

Result🔗

All twelve turn files were unique and passed the frozen mechanical duration checks. The isolated and conversation-aware assemblies were 12.61 and 13.25 seconds respectively.

Before reveal, the isolated render passed only the turn-timing gate. Its voices were not judged natural, the speakers did not sound like they were reacting to each other, and the complete exchange was not convincing. The conversation-aware render passed all four gates: its voices sounded natural, its turns sounded mutually responsive, its timing felt conversational, and the complete scene was judged convincing.

This is positive evidence that broad, scene-grounded performance direction can produce a convincing short two-speaker render with the current local apparatus. It is not evidence that dialogue context reliably causes the improvement: the pilot contains one scene, one sample per arm, and one listener, and its authored directions differ turn by turn. A replication should freeze multiple scenes and seeds, preserve blind whole-conversation review, and include an ablation separating shared scene context from per-turn performance direction.