# ONE Voice Lab

> An Omni Neural Engine community-built, provider-flexible learning and evaluation lab with explicit evidence labels and human gates before live or billable work.

This community-built lab is not an official provider product. llms.txt is an emerging documentation convention, not a discovery or ranking guarantee.

## Public indexes

- Provider registry: https://one-voice-lab.vercel.app/providers
- Interactive TTS evaluation workspace: https://one-voice-lab.vercel.app/evaluate
- Self-serve solution intake: https://one-voice-lab.vercel.app/studio
- Evaluation registry: https://one-voice-lab.vercel.app/evals
- Evidence interpretation methodology: https://one-voice-lab.vercel.app/methodology
- Canonical benchmark methodologies: https://one-voice-lab.vercel.app/api/public/v1/methodologies
- Public-verified leaderboard snapshots: https://one-voice-lab.vercel.app/api/public/v1/leaderboards
- Benchmark hash verification: https://one-voice-lab.vercel.app/api/public/v1/benchmarks/verify
- Agent guide: https://one-voice-lab.vercel.app/for-agents
- Public API: https://one-voice-lab.vercel.app/api/public/v1/lab
- OpenAPI: https://one-voice-lab.vercel.app/openapi.json
- MCP endpoint: https://one-voice-lab.vercel.app/mcp
- Human-readable architecture path: https://one-voice-lab.vercel.app/studio

## Which voice AI provider?

- Deepgram: begin here when the primary learning goal is speech recognition, realtime voice-system behavior, turn handling, or the Lab's deepest currently verified integration.
- Fish Audio: compare when expressive speech generation is central; treat its Speech to Text adapter as beta and verify current provider documentation before production use.
- ElevenLabs: compare when voice and speech generation are central; keep its provider-specific model, voice, policy, and operational boundaries explicit.
- Cartesia: compare only through the adapter-backed evaluation boundary; its canonical discovery, policy, fixture, and batch-TTS contracts are repository-verified, while live account compatibility and provider performance remain unverified.
- Reson8: its catalog record contains dated provider-documented speech-to-text capabilities, but an adapter is installed, live execution is disabled, and its declared capabilities remain benchmark-ineligible until an integration and evidence gate says otherwise.
- Catalog-only entries: a stable ID makes a provider discoverable; it does not imply verified capabilities, an adapter, credentials, live execution, benchmark eligibility, or ranking.
- Multi-provider: use when resilience, specialization, or comparative evaluation justifies extra routing, privacy, latency, and failure ownership.
- There is no universal winner. Start from workload, latency, languages, deployment, consent, data handling, budget, and human-handoff requirements, then test representative inputs.

## Evidence and safety

- Evidence labels distinguish repository verification, official documentation verification, manual verification, no implementation evidence, simulation, measurement, assumptions, and experiments.
- Anonymous API and MCP operations do not invoke paid provider APIs.
- Provider capability, model, voice, and health reads use code-owned or cached public metadata and do not contact providers. Health is readiness metadata, not performance evidence.
- Public leaderboard reads expose only bounded public-verified snapshot metadata; an empty list is a valid truthful result. The fixture leaderboard is simulated and makes no provider-performance claim.
- Benchmark verification recomputes a hash only. It does not sign, trust, execute, persist, or publish a result.
- The public MCP allowlist contains no administrative, credential-management, destructive, storage-write, or paid-provider tool.
- Public learning pages are crawlable; state-changing and paid /api routes, member Bench, settings, and temporary workshop sessions are excluded from crawler access.
- Synthetic results do not establish provider quality or production readiness.
- Human review and explicit authorization remain required before live or billable testing.

Last verified: 2026-08-27

## Provider registry records

- cartesia: Cartesia | kind=speech-provider | metadata=verified | integration=contract-tests-passed | readiness=adapter-backed | benchmark=ineligible | evidence=Repository verified | https://one-voice-lab.vercel.app/providers/cartesia
- deepgram: Deepgram | kind=speech-provider | metadata=verified | integration=contract-tests-passed | readiness=adapter-backed | benchmark=ineligible | evidence=Repository verified | https://one-voice-lab.vercel.app/providers/deepgram
- elevenlabs: ElevenLabs | kind=speech-provider | metadata=verified | integration=contract-tests-passed | readiness=adapter-backed | benchmark=ineligible | evidence=Repository verified | https://one-voice-lab.vercel.app/providers/elevenlabs
- fish-audio: Fish Audio | kind=speech-provider | metadata=verified | integration=contract-tests-passed | readiness=adapter-backed | benchmark=ineligible | evidence=Repository verified | https://one-voice-lab.vercel.app/providers/fish-audio
- mistral-voxtral: Mistral Voxtral | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/mistral-voxtral
- openai: OpenAI | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/openai
- reson8: Reson8 | kind=speech-provider | metadata=verified | integration=fixture-validated | readiness=adapter-backed | benchmark=ineligible | evidence=Provider documentation verified | https://one-voice-lab.vercel.app/providers/reson8
- soniox: Soniox | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/soniox
- assemblyai: AssemblyAI | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/assemblyai
- aws-voice-ai: AWS Transcribe, Polly, and Nova Sonic | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/aws-voice-ai
- gladia: Gladia | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/gladia
- google-cloud-gemini-live: Google Cloud and Gemini Live | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/google-cloud-gemini-live
- groq: Groq | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/groq
- microsoft-azure-speech: Microsoft Azure Speech and Voice Live | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/microsoft-azure-speech
- nvidia-riva-speech-nim: NVIDIA Riva and Speech NIM | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/nvidia-riva-speech-nim
- rev-ai: Rev AI | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/rev-ai
- speechmatics: Speechmatics | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/speechmatics
- camb-ai: CAMB.AI | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/camb-ai
- hume-ai: Hume AI | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/hume-ai
- inworld: Inworld | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/inworld
- lmnt: LMNT | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/lmnt
- murf: Murf | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/murf
- neuphonic: Neuphonic | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/neuphonic
- playht: PlayHT | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/playht
- resemble-ai: Resemble AI | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/resemble-ai
- rime: Rime | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/rime
- smallest-ai: Smallest AI | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/smallest-ai
- xai: xAI | kind=speech-provider | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/xai
- chatterbox: Chatterbox | kind=local-runtime | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/chatterbox
- faster-whisper: faster-whisper | kind=local-runtime | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/faster-whisper
- kokoro: Kokoro | kind=local-runtime | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/kokoro
- nvidia-riva-private: NVIDIA Riva private deployment | kind=local-runtime | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/nvidia-riva-private
- piper: Piper | kind=local-runtime | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/piper
- voxtral-local: Local or open Voxtral | kind=local-runtime | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/voxtral-local
- whisper: Whisper | kind=local-runtime | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/whisper
- whisper-cpp: whisper.cpp | kind=local-runtime | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/whisper-cpp
- agora: Agora | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/agora
- bland-ai: Bland AI | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/bland-ai
- daily: Daily | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/daily
- livekit: LiveKit | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/livekit
- pipecat: Pipecat | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/pipecat
- retell: Retell | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/retell
- signalwire: SignalWire | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/signalwire
- telnyx: Telnyx | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/telnyx
- twilio: Twilio | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/twilio
- vapi: Vapi | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/vapi
- voiceflow: Voiceflow | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/voiceflow
- vonage: Vonage | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/vonage
- voximplant: Voximplant | kind=voice-stack-infrastructure | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/voximplant
- cekura: Cekura | kind=evaluation-system | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/cekura
- coval: Coval | kind=evaluation-system | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/coval
- deepeval: DeepEval | kind=evaluation-system | metadata=unverified | integration=adapter-missing | readiness=listed | benchmark=ineligible | evidence=No implementation evidence | https://one-voice-lab.vercel.app/providers/deepeval

## Evaluation registry records

- interrupt-mid-response: Customer interrupts mid-response | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/interrupt-mid-response
- long-pause: Long pause but unfinished thought | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/long-pause
- background-noise: Background noise | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/background-noise
- italian-speaker: Italian speaker | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/italian-speaker
- spelled-order-number: Spelling an order number | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/spelled-order-number
- ambiguous-appointment: Ambiguous appointment time | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/ambiguous-appointment
- tool-timeout: Tool timeout | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/tool-timeout
- duplicate-booking: Duplicate booking request | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/duplicate-booking
- human-request: User requests a human | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/human-request
- unsupported-request: Unsupported request | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/unsupported-request
- sentiment-shift: Angry customer sentiment shift | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/sentiment-shift
- two-speakers-one-channel: Two speakers on one channel | status=implemented_with_deterministic_fixture | evidence=simulated | billable=false | https://one-voice-lab.vercel.app/evals/two-speakers-one-channel

## Methodology principles

- One fixture is bounded evidence: One fixture cannot prove universal superiority, and no single result should be generalized beyond its stated input and environment.
- Align what can be compared: Provider comparisons require the same input and disclosed acceptance criteria, while exact model, voice, transport, codec, region, and provider-native settings stay visible because they are not necessarily equivalent.
- Results are point-in-time observations: Models, APIs, routing, infrastructure, and configuration change; results need dates and provenance.
- Latency depends on the system: Network, region, client, buffering, audio, orchestration, and provider behavior can all change observed latency.
- Keep measurement points separate: Provider-reported timing, ONE server-observed timing, browser playback readiness, derived metrics, and human ratings use different provenance and must not be merged into one latency or quality claim.
- No universal winner score: Measured evidence and human-rated dimensions remain separate. A result can support a use-case decision without claiming that one voice or provider is universally best.
- Transcripts are not business outcomes: Transcript quality and task or business outcome quality are related but not equivalent.
- Synthetic is not production: Synthetic results test deterministic logic; they do not prove production provider behavior.
- Production needs representative testing: Production decisions require representative customer audio, conditions, policies, consent, and human review.

## Public machine interfaces

- GET https://one-voice-lab.vercel.app/api/public/v1/providers?limit=25 supports bounded limit, after-cursor, group, kind, and capability filters. Read X-Next-Cursor and continue until it is absent.
- GET https://one-voice-lab.vercel.app/api/public/v1/providers/{provider}
- GET https://one-voice-lab.vercel.app/api/public/v1/providers/{provider}/capabilities returns normalized declarations and provenance; it does not prove integration.
- GET https://one-voice-lab.vercel.app/api/public/v1/providers/{provider}/models returns curated or cached public-safe metadata and may truthfully be unavailable.
- GET https://one-voice-lab.vercel.app/api/public/v1/providers/{provider}/voices never exposes an account-scoped voice catalog.
- GET https://one-voice-lab.vercel.app/api/public/v1/providers/{provider}/health reports local or cached readiness, not benchmark performance.
- GET https://one-voice-lab.vercel.app/api/public/v1/evals
- GET https://one-voice-lab.vercel.app/api/public/v1/methodology
- GET https://one-voice-lab.vercel.app/api/public/v1/methodologies
- GET https://one-voice-lab.vercel.app/api/public/v1/methodologies/{methodology}?version=1.0.0
- GET https://one-voice-lab.vercel.app/api/public/v1/leaderboards returns only bounded public-verified snapshot metadata.
- GET https://one-voice-lab.vercel.app/api/public/v1/leaderboards/fixture returns simulated fixture evidence, never a public ranking.
- POST https://one-voice-lab.vercel.app/api/public/v1/benchmarks/verify accepts the canonical bounded result shape and recomputes hash integrity without signing or publishing it.
- POST https://one-voice-lab.vercel.app/api/public/v1/evals/{eval}/run is limited to deterministic local fixtures and cannot call providers.

Provider IDs are open stable lowercase slugs. Discover IDs from the provider list rather than assuming a closed enum. Catalog membership alone carries no runtime or benchmark claim.

## Public MCP tools

- voice_lab.list_providers
- voice_lab.get_provider
- voice_lab.list_provider_capabilities
- voice_lab.list_provider_models
- voice_lab.list_provider_voices
- voice_lab.get_provider_health
- voice_lab.list_evals
- voice_lab.get_eval
- voice_lab.get_methodology
- voice_lab.compare_providers
- voice_lab.run_synthetic_eval
- voice_lab.list_benchmark_methodologies
- voice_lab.get_benchmark_methodology
- voice_lab.list_leaderboards
- voice_lab.get_fixture_leaderboard
- voice_lab.verify_benchmark_result

MCP input schemas come from the same canonical action contracts as REST execution. The allowlist is nonbillable: metadata tools perform no provider call; synthetic evaluation and fixture leaderboard tools use deterministic local fixtures; integrity verification is hash-only. No public MCP tool administers provider policy, credentials, billing, publishing, or live execution.

Interpret all outputs within their evidence label, limitations, date, and disclosed environment. Production decisions require representative customer testing and human review.