Build on the supported package contract
Use @absolutejs/voice-benchmarks through its supported public entry points.
@absolutejs/voice-benchmarksv0.0.0betaDev ToolsPerformance & accuracy benchmarks for @absolutejs/voice (STT, TTS, duplex, telephony, sessions) vs. Vapi.
# Private workspace project: ~/abs/benchmarks/voicePerformance & accuracy benchmarks for @absolutejs/voice — STT, TTS, duplex, telephony, and session soak tests, plus head-to-head comparisons against Vapi.
These consume the published @absolutejs/voice package and provider adapters, so they double as a real-world consumer of the public API.
For intake transcription, run bun run bench:intake:accuracy. It compares Deepgram Flux with OpenAI on the domain-heavy fixture set and reports both WER and expected-term recall, so a provider cannot pass by getting filler words right while missing the names and phrases that populate durable profile fields.
Benchmarks are exposed as bench: scripts. Examples:
Run bun run to list every benchmark target.
Results are written to benchmark-results/ (gitignored). Some accent/code-switch suites read the multilingual corpus from a sibling voice-fixtures-multilingual checkout (../../voice-fixtures-multilingual).
benchmarks/vapi-baseline.example.json is a template — copy it to benchmarks/your-vapi-metrics.json (gitignored) with your own Vapi numbers to compare against.
Outcomes
Use @absolutejs/voice-benchmarks through its supported public entry points.
Hardening checklist
Follow in order
Working example for Setup.
bun install
cp .env.example .env # add provider API keys for live benchmarksBenchmarks are exposed as bench: scripts. Examples:
bun run bench:tts # TTS latency across providers
bun run bench:stt all telephony # STT accuracy on telephony fixtures
bun run bench:duplex # full-duplex turn-taking
bun run bench:telephony:run # telephony suite -> benchmark-results/