Speech to Text (Beta) | Murf API

Murf Speech to Text (STT) converts live audio into text in real time. You stream audio over a WebSocket and receive partial transcripts while the caller is speaking, then a final transcript with word-level timestamps when the turn ends. It is built for voice agents, IVR, and call analytics, and accepts telephony audio (8 kHz µ-law and A-law) directly.

Combined with Falcon 2.1, Murf now covers both ends of a voice agent: hearing the caller and speaking back.

Ideal for

Voice agents and IVR: Show the caller’s words as they speak, and pass each completed turn to your LLM. The server ends each turn when the caller pauses.
Collections and payment reminders: Boost recognition of product names and account terms with keyword boosting.
Customer support: Transcribe calls in the caller’s own language, or detect it automatically.
Call analytics and QA: Get word-level timestamps on each final transcript for search, redaction, and review.

Models

	Finch-Indic	Finch-Global
Best for	India: Indic languages, hosted in India	US, UK, Australia, and Europe: English and European languages, hosted in US East
`model`	`finch_indic_1.0`	`finch_global_1.0`
Host	`in.api.murf.ai`	`us-east.api.murf.ai`
Languages	English (India), Hindi, Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu	English (US, UK, Australia), German, Spanish, French, Italian, Portuguese and Dutch
Detection	Omit `language` or set `auto`. Detects only among this model’s languages.	Omit `language` or set `auto`. Detects only among this model’s languages.

Every language code is listed on Languages.

Key features

Real-time streaming

Stream audio over one WebSocket. partial transcripts update while the caller speaks, and a final transcript arrives when the turn ends.

Languages and auto-detect

Both models transcribe multiple languages, and can detect the language from the audio. See Languages.

Background Voice Cancellation

Set bvc=true when you connect to remove background voices before transcription.

Telephony-ready audio

Send 8 kHz µ-law or A-law straight from your telephony provider, plus PCM and Opus at rates up to 48 kHz.

Keyword boosting

Improve recognition of brand names, product terms and domain vocabulary per connection.

Multi-turn sessions

Keep one connection open for a whole call. Turns end when the caller pauses; timestamps stay on one session timeline.

End-of-turn detection

The server ends each turn when the caller pauses.

Word-level timestamps

Every final transcript includes start and end times for each word.

How it works

Connect

Open a WebSocket to the host for your model: wss://in.api.murf.ai/v1/transcription/stream-input for Finch-Indic, or wss://us-east.api.murf.ai/v1/transcription/stream-input for Finch-Global. Pass model and your API key, and optionally language and keywords.

Stream audio

Send audio as binary frames. The server detects speech and returns partial messages as the turn grows. The default format is 16-bit PCM, 16 kHz, mono; for telephony or other formats, see Audio formats.

End turns and the session

The server ends each turn when the caller pauses and sends its final transcript. Send {"type": "Finalize"} only if you want to end a turn before the next pause. Send {"type": "CloseStream"} when the call ends.

Pricing

Speech to Text costs $0.12 per hour of audio, billed per second. Enterprise customers can get custom pricing; contact sales.

See Pricing

Get started

Streaming

Transcribe a file or your microphone in a few minutes.

API reference

Every message, field and close code.

Audio formats

Codecs, sample rates and telephony examples.

Languages

Supported languages and automatic detection.