October 22, 2026
By Steven Van ·
Murf's beta WebSocket service covers 12 languages including Hindi and Tamil, supports 8 kHz telephony audio, and costs $0.24 per hour.
Murf has put Murf AI's Speech to Text into beta. It transcribes audio in real time over a single WebSocket connection and is aimed at voice agents, IVR and call analytics. It is available in the India region only.
Languages and transcripts
It supports 12 languages: English (India), Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu. Automatic language detection is also available. Language, keywords and audio format are set when connecting, so changing any of them means opening a new connection.
The server sends partial messages while the caller speaks. Each one contains the whole turn so far, so a client should replace the displayed text rather than append to it. A final message follows when the turn ends, with word-level timestamps measured from the start of the session. Keyword boosting is available for brand names and domain terms.
Turns and sessions
The server does not end a turn on silence. A turn ends only when the client sends Finalize or CloseStream. After Finalize, the connection stays open and further audio starts the next turn, so one connection can cover a whole call. While the agent is speaking, a KeepAlive message stops the connection closing for inactivity. Automatic end-of-turn detection (commit_mode=auto) and background voice removal (bvc=true) are listed as coming soon and currently have no effect.
Audio formats
Telephony audio can be sent as it is, as 8 kHz µ-law or A-law. PCM and Opus are supported up to 48 kHz. Stereo input is mixed down to mono, and speaker diarization is not available. There is no separate batch endpoint, so recorded files are transcribed by streaming them over the WebSocket.
Pricing and limits
- $0.24 per hour of audio, billed per second.
- Free tier: 5 concurrent requests.
- Pay as you go: 20 concurrent by default, up to 150 on request.
- Enterprise: custom limits (150+) and custom pricing.
Each open connection uses one slot until it closes. These limits are separate from Text to Speech concurrency.
Concurrency reporting on Falcon 2.1 streaming
Falcon 2.1 streaming responses now report the serving region and current concurrency usage. HTTP streaming adds the headers X-Murf-Region, X-Murf-Concurrency-Limit, X-Murf-Concurrency-Used and X-Murf-Concurrency-Remaining. WebSocket final messages include a concurrency object, and connecting with include_usage=true also returns session_info when a context starts.