Question 1

How is speech-to-speech interpretation different from machine translation?

Accepted Answer

Machine translation optimizes for literal word-for-word accuracy at any latency cost. Interpretation mimics human behavior: it removes accidental repetition, compresses rambling phrasing, softens gratuitous expletives, replaces figurative language with clear meaning, and prioritizes natural spoken delivery under time pressure. Babelbit scores based on semantic meaning (not exact text match), speaking rate plausibility, and latency—not literal translation accuracy.

Question 2

What do validators score miners on?

Accepted Answer

Validators transcribe miner audio using their own STT model (faster-whisper), embed both miner transcript and reference text using sentence transformers, and check three gates: (1) semantic accuracy ≥0.65; (2) speaking rate between 0.3x and 1.3x reference speed; (3) latency within 30% of source duration (min 2s, max 10s). Audio passing the first two gates is scored on latency; audio failing either gate scores zero.

Question 3

How much compute do validators need?

Accepted Answer

Based on CPU-only benchmarks, the minimum is 8 vCPU / 16 GB RAM. A 250-miner, 60-second synthetic batch took about 23–24 minutes on an 8 vCPU AMD EPYC system but hours on a 4 vCPU Intel system. STT transcription is the bottleneck. Validators should not run below 8 vCPU / 16 GB in production; anything smaller will push scoring into unacceptable time windows.

Question 4

Can miners use traditional STT-translate-TTS pipelines?

Accepted Answer

Yes. The validator does not enforce a specific architecture. Miners can use cascaded STT-translate-TTS, speech-token models (like Moshi), recurrent architectures (Mamba, RWKV), or language-model fine-tuning. What matters is validator-facing behavior: miners must accept signed Bittensor requests, return audio, and achieve low latency while preserving meaning and sounding natural.

Question 5

What is the qualifying round versus The Arena?

Accepted Answer

The qualifying round distributes 20% of emissions proportionally to all miners based on their challenge scores—most will not get rich but receive rewards for hard work and improvement. Top qualifiers compete in The Arena for the remaining 80% on a winner-takes-all basis. Arena eligibility is derived from recent qualifying performance.

Question 6

How do miners get discovered and scored?

Accepted Answer

Qualifying miners are discovered from Bittensor axon metadata alone (IP, port, registration on netuid 59). The validator discovers them, generates challenges, sends signed requests, collects audio responses, and scores them. Managed submission flows (arena) also require a Docker image and Hugging Face repository handle so the validator can manage runtime and model version.

Subnet 59

Subnet 59

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup