7 speech models
Speech & audio models
Transcription bills per minute of audio and synthesis bills per character, so neither compares cleanly to token pricing. This table keeps each model in its own native unit rather than forcing a misleading conversion.
Pricing checked against provider documentation on . How we verify
7 of 7 models
| Model | Provider | Price | Context | Best for | Status |
|---|---|---|---|---|---|
| Universal-2 | AssemblyAI | $0.0025 / min | — | Cheapest broad-language transcription | stable |
| GPT-4o Mini Transcribe | OpenAI | $0.003 / min | — | Low-cost transcription inside the OpenAI stack | stable |
| Scribe v2 | ElevenLabs | $0.0037 / min | — | Widest language coverage with diarization | stable |
| Nova-3 | Deepgram | $0.0043 / min | — | Real-time streaming transcription | stable |
| GPT-4o Transcribe | OpenAI | $0.006 / min | — | Higher-accuracy managed transcription | stable |
| Flash / Turbo TTS | ElevenLabs | $50.00 / 1M ch | — | Low-latency speech synthesis | stable |
| Multilingual v3 TTS | ElevenLabs | $100.00 / 1M ch | — | Highest-quality voice output | stable |
No models match those filters.
Accuracy has largely converged
At the top end, word error rates on clean audio differ by fractions of a percent. Differentiation now comes from diarization quality, latency, and language coverage — not headline accuracy.
Streaming changes the maths
Real-time transcription costs more per minute than batch and usually caps concurrency. If your product can tolerate a few seconds of delay, batch endpoints are substantially cheaper.
Speech-to-speech is its own category
Realtime voice APIs bill audio input and output as tokens rather than minutes, which makes them look expensive next to transcription. They replace a whole STT → LLM → TTS pipeline, so compare totals.