Speech model performance
Latency, accuracy, and cost across STT, TTS, and agentic models, over HTTP and WebSocket. Methodology in the open.
STT · 20 samples per model · LibriTTS-R · run 2026-10-29
Lower is better. Every successful sample is plotted. Batch is timed from the start of the request to the full transcript. Streaming also records when the first text arrived. Both start the clock at the request, so they share an axis.
Latency vs accuracy
Transcript complete p50 on x, WER (mean) on y. Bottom-left is fastest and most accurate.
Latency vs cost
Transcript complete p50 on x, published price per audio hour on y, using fixed conversion assumptions. Bottom-left is fastest and cheapest.
| Provider / Model | First text p50 | Transcript complete p50 | Transcript complete p95 | WER | CER | $/audio hr |
|---|---|---|---|---|---|---|
Sprag RhythmHTTP streaming | 189 | 194 | 538 | 2.2% | 0.6% | $0.075 |
Sprag RhythmHTTP batch | n/a | 244 | 484 | 2.2% | 0.6% | $0.075 |
Sprag SymphonyHTTP batch | n/a | 264 | 659 | 1.7% | 0.4% | $0.09 |
Cartesia Ink WhisperHTTP batch | n/a | 289 | 774 | 2.2% | 0.7% | $0.0673 |
Deepgram Nova 3HTTP batch · 14/20 runs | n/a | 337 | 535 | 2.4% | 0.7% | $0.258 |
Sprag SymphonyHTTP streaming | 260 | 338 | 708 | 2.2% | 0.6% | $0.09 |
ElevenLabs Scribe v2HTTP batch | n/a | 436 | 930 | 1.6% | 0.4% | $0.22 |
OpenAI GPT TranscribeHTTP batch | n/a | 714 | 1,439 | 1.9% | 0.5% | $0.27 |
OpenAI GPT TranscribeHTTP streaming | 608 | 792 | 1,591 | 1.9% | 0.5% | $0.27 |
Soniox STT Async v5HTTP batch | n/a | 2,000 | 2,272 | 2.1% | 0.6% | $0.0975 |
AssemblyAI Universal-3.5 ProHTTP batch | n/a | 2,485 | 4,013 | 1.7% | 0.5% | $0.21 |
Price sources and conversion assumptions
Regular published list rates, normalized with fixed assumptions. Prices do not depend on measured recording lengths or token usage. Promotions, free credits, taxes and enterprise discounts are excluded.
- One input audio hour. For token-priced transcript output, use the published reference of 15,000 text tokens/hour; no optional context or add-ons.
- Scale monthly conversion: $299 / 8,000,000 credits; source https://www.cartesia.ai/pricing.
- AssemblyAI / Universal-3.5 Pro · HTTP batch: $0.21 / input audio hour. Published source · checked 2026-10-29
- Cartesia / Ink Whisper · HTTP batch: 0.5 credits / input audio second; Scale: $299 / 8M credits monthly. Published source · checked 2026-10-29
- Deepgram / Nova 3 · HTTP batch: $0.0043 / input audio minute (monolingual PAYG list). Published source · checked 2026-10-29
- ElevenLabs / Scribe v2 · HTTP batch: $0.22 / input audio hour. Published source · checked 2026-10-29
- OpenAI / GPT Transcribe · HTTP batch: $0.0045 / input audio minute. Published source · checked 2026-10-29
- Soniox / STT Async v5 · HTTP batch: $1.5 / 1M input audio tokens + $3.5 / 1M input/output text tokens. Published source · checked 2026-10-29
- Sprag / Rhythm · HTTP batch: $0.075 / input audio hour. Published source · checked 2026-10-29
- Sprag / Symphony · HTTP batch: $0.09 / input audio hour. Published source · checked 2026-10-29
- OpenAI / GPT Transcribe · HTTP streaming: $0.0045 / input audio minute. Published source · checked 2026-10-29
- Sprag / Rhythm · HTTP streaming: $0.075 / input audio hour. Published source · checked 2026-10-29
- Sprag / Symphony · HTTP streaming: $0.09 / input audio hour. Published source · checked 2026-10-29
- Conversion reference: 60,000 characters and approximately 15,000 text tokens per speech hour; Sprag audio token = 100 ms.
- Conversion reference: Approximately 0.3 text tokens per character and 30,000 audio tokens per speech hour.
- Conversion reference: Input audio token per 100 ms; output audio token per 50 ms.