Realtime speech model pricing compared
Two kinds of “live”: streaming transcription for captions, and speech-to-speech models for voice chat. Both normalised to cost per hour.
Prices checked 2026-10-26. These are upstream list prices, not BibiGPT pricing.
Live transcription
| Model | per hour | 10 h | List price | Strengths | In BibiGPT |
|---|---|---|---|---|---|
Apple Speech (on-device, macOS) Apple SFSpeechRecognizerFree and works offline; less accurate than cloud models. | $0 | $0 | Free (built into the system) | System languages Low latency | On device Mac app, live mic dictation |
Qwen-Audio-3.1 ASR StreamingCheapest Alibaba Qwen qwen-audio-3.1-asr-flash-streamingVendor claims ~160 ms to first token (not measured). An order of magnitude above the file version. | $0.07Est. | $0.7 | ¥6 input / ¥4.5 output per 1M tokens (Beijing region) | 30 languages + 10 dialects Low latencyHotwordsDialects | Not in BibiGPT |
Soniox Real-time Soniox stt-rtThe cheapest realtime transcription we found outside China; can translate as it goes. | $0.12 | $1.2 | $0.12/hour | 60 languages Low latencySpeaker labelsWord timestamps | Not in BibiGPT |
Paraformer Realtime v2 Alibaba Qwen paraformer-realtime-v2The cheapest and most predictable realtime ASR in China. | $0.122 | $1.22 | ¥0.00024/second (10 free hours a month) | zh/en/ja/ko + accents Low latencyHotwordsDialects | Not in BibiGPT |
Doubao streaming ASR 2.0 ByteDance Volcengine Doubao streaming ASR model 2.0Strong realtime Chinese at about the Paraformer Realtime price. | $0.141 | $1.41 | ¥1/hour (per mono channel, tiered) | zh/en + dialects Low latencyHotwordsDialects | Not in BibiGPT |
AssemblyAI Universal-Streaming AssemblyAI universal-streamingA common ears-only layer for voice agents. The Pro Realtime tier is $0.45/h. | $0.15 | $1.5 | $0.15/hour | English + multilingual Low latencyWord timestamps | Not in BibiGPT |
Qwen3-ASR Flash Realtime Alibaba Qwen qwen3-asr-flash-realtimeRealtime ASR from China with emotion labels. | $0.167 | $1.67 | ¥0.00033/second (same as fun-asr-realtime) | ~27 languages + 4 dialects Low latencyEmotionDialects | Not in BibiGPT |
Grok Voice Transcribe (streaming) xAI grok-voice-transcribe-2.0Cheap live captions for English; no Chinese. | $0.2 | $2 | $0.20/hour (WebSocket) | 25 languages (no CJK) Low latencyWord timestampsSpeaker labels | Not in BibiGPT |
Deepgram Nova-3 Streaming Deepgram nova-3A long-standing pick for English calls and meetings. Current price is a limited-time promo. | $0.288 | $2.88 | $0.0048/minute (monolingual, promo price; multilingual $0.0058) | Multilingual, strongest in English Low latencySpeaker labelsHotwords | Not in BibiGPT |
Fish Audio Transcribe Stream Fish Audio transcribe-stream-1Billed per session hour, silence included. | $0.36 | $3.6 | $0.36/session hour (preview) | Multilingual Low latency | Not in BibiGPT |
ElevenLabs Scribe v2 Realtime ElevenLabs scribe_v2_realtimeMultilingual live captions, ~150 ms latency (vendor figure). | $0.39 | $3.9 | $0.39/hour | 90+ languages Low latencyWord timestamps | Not in BibiGPT |
Gemini 3.5 Transcribe Live Google gemini-3.5-transcribe-liveAbout 80% above the file version, in exchange for text while you speak. | $0.542Est. | $5.42 | $3.50 audio input / $21 text output per 1M tokens (official estimate about $0.009/minute) | 85+ languages Low latencySpeaker labelsFiller removal | In BibiGPT Live transcription (web / desktop) |
GPT Live Transcribe OpenAI gpt-live-transcribeOn the pricey end for live captions; gpt-realtime-translate is $0.034/min. | $1.02 | $10.2 | $0.017/minute | Multilingual Low latency | Not in BibiGPT |
Realtime voice conversation
| Model | per hour | 10 h | List price | Strengths | In BibiGPT |
|---|---|---|---|---|---|
Qwen-Audio-3.1 Realtime Plus Alibaba Qwen qwen-audio-3.1-realtime-plusSpeech-to-speech reachable from mainland China. Same list price as 3.0; context is re-billed every turn. | $0.602Est. | $6.02 | ¥40 audio input / ¥150 audio output per 1M tokens (Beijing region) | Chinese and English first Speech-to-speechLow latency | Not in BibiGPT |
Gemini 3.1 Flash Live Google gemini-3.1-flash-live-previewNative audio in and out, works with screen sharing. The cheapest speech-to-speech option here. | $0.675Est. | $6.75 | $3 audio input / $12 audio output per 1M tokens (official estimate $0.005 / $0.018 per minute) | Multilingual Speech-to-speechLow latency | In BibiGPT Gemini Desktop & web Live voice chat |
GPT Realtime Mini OpenAI gpt-realtime-miniThe budget GPT Realtime, about a third of the full model. | $0.9Est. | $9 | $10 audio input / $20 audio output per 1M tokens | Multilingual Speech-to-speechLow latency | Not in BibiGPT |
GPT Realtime OpenAI gpt-realtimeThe common baseline for voice agents; long sessions also pay for context tokens. | $2.88Est. | $28.8 | $32 audio input / $64 audio output per 1M tokens (gpt-realtime-2 audio is the same) | Multilingual Speech-to-speechLow latency | Not in BibiGPT |
Grok Voice 3.0 xAI grok-voice-3.0Flat per-minute price: easy to budget, highest rate here. | $4.8 | $48 | $0.08/minute (plus text input) | Multilingual Speech-to-speechLow latency | Not in BibiGPT |
- Gemini audio input: 25 tokens per second; transcript output about 3 tokens per second.
- Qwen-Audio-3.1: about 20 input + 4 output tokens per second, measured on our own Chinese samples.
- Voice conversation assumes the user talks 50% of the hour and the model the rest.
- CNY ↔ USD at 7.1.
FAQ
Which transcription model is cheapest?
On our samples Qwen-Audio-3.1 file transcription came to about ¥0.075–0.10 per hour, below StepAudio 2.5 (¥0.15/h) and Paraformer v2 (¥0.288/h). On-device models in the Bibi Local desktop app cost nothing.
What is the difference between STT, live and TTS?
STT turns a finished recording into text and is billed per audio hour. Live models do the same while you speak, or answer you in voice, and cost more. TTS goes the other way: text in, speech out, usually billed per character.
Why are some prices marked “Est.”?
Those models bill by tokens, not minutes. We convert with the token densities listed under the table; your real bill moves with speech speed and language.
Can I use these models in BibiGPT?
Models marked “In BibiGPT” are covered by membership. “Your key” models are selectable once you paste your own API key. “On device” models run free in the Bibi Local desktop app. “Not in BibiGPT” rows are listed for reference.
Try them on your own audio
Paste a YouTube, Bilibili or podcast link into BibiGPT and pick the engine in the transcription settings. For private recordings, Bibi Local transcribes on your own computer.