Changelog

Speech to Text now requires model on the WebSocket URL.

• Finch-Indic (finch_indic_1.0) transcribes English (India) and 11 Indian languages at wss://in.api.murf.ai/v1/transcription/stream-input
• Finch-Global (finch_global_1.0) transcribes English (US, UK, Australia), German, Spanish, French, Italian, Portuguese and Dutch at wss://us-east.api.murf.ai/v1/transcription/stream-input

Speech to Text can remove background voices before transcription. Pass bvc=true when you connect; it is off until you set it. Available on Finch-Indic and Finch-Global.

See the Streaming guide.

List Voices takes model. Pass falcon-2.1 to list voices for streaming, or gen2 to list voices for synthesis.

Each model has its own catalog. Fetch the matching list before you send voiceId on Stream Speech, WebSockets, or Synthesize Speech.

See the Voice Library for both catalogs.

Murf Speech to Text (Beta) is now available: real-time transcription over WebSocket for voice agents, IVR and call analytics.

• 12 languages: English (India), Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu, plus automatic language detection
• Partial and final transcripts: partial updates while the caller speaks; a final transcript with word-level timestamps when you end the turn
• Telephony-ready: send 8 kHz µ-law or A-law directly, as well as PCM and Opus up to 48 kHz
• Keyword boosting for brand names and domain terms
• Multi-turn sessions: end each turn with Finalize and keep one connection for the whole call
• $0.24 per hour, billed per second. Free tier: 5 concurrent requests. Pay as you go: Up to 150 concurrent requests (20 concurrent as default ; Increasable based on discussion)

Falcon 2.1 streaming responses now report the serving region and your current concurrency usage.

• HTTP streaming includes X-Murf-Region, X-Murf-Concurrency-Limit, X-Murf-Concurrency-Used, and X-Murf-Concurrency-Remaining.
• WebSocket final messages include a concurrency object. Connect with include_usage=true to also receive session_info when a context starts.

Falcon 2.1 is now live on Stream Speech and WebSockets.

• More natural voice quality across languages and voices.
• Lower time to first spoken word, as benchmarked independently by Coval.
• Fewer artifacts and hallucinations, with more consistent loudness between responses.
• Exact pronunciation control through inline IPA or a pronunciation dictionary, on both Stream Speech and WebSockets.
• More languages and voices in the Voice Library.

Falcon 2 now gives you two ways to set exactly how a word is pronounced:

• IPA in your text: write the word in IPA between forward slashes inside double quotes, for example Welcome to "/mɜːrf/". Works on Stream Speech and WebSockets.
• Pronunciation dictionary: pass pronunciationDictionary on Stream Speech, the same parameter Synthesize Speech supports for Gen2. On WebSockets, send pronunciation_dictionary in voice_config once per context. Each entry sets a word’s IPA or replacement words, for every time the word appears.

Falcon 2 streaming now supports volumeOffset on Stream Speech and in WebSocket voice_config. Adjust loudness relative to the model’s default output (integer -50 to 50, default 0) so voice agents match phone lines, earpieces, and noisy environments without adding latency.

See Volume offset (Falcon 2) and the Falcon 2 model page for suggested ranges and best practices.

You can now request per-word timing on the WebSocket API with Falcon 2. Set add_word_timestamps to true; once enabled, it applies to the whole WebSocket connection. Falcon returns each spoken word with start_s and end_s in separate JSON frames alongside the audio.

This is especially useful for voice agents handling barge-in: you can see how much of a response had already played when the user interrupts.

The Gen2 model has been decommissioned for the Streaming API. If you haven’t migrated yet, use our latest Falcon 2 model on the Streaming API.

The Gen2 model remains available for non-streaming text-to-speech through the Synthesize Speech endpoint.

Our legacy Gen2 streaming model is set to be deprecated on 16th August 2026. We recommend migrating to our latest Falcon 2 model promptly.

Gen2 non-streaming text-to-speech will remain available through the Synthesize Speech endpoint.

We’re excited to announce a major release: the new Falcon 2 model is now live, delivering significant improvements in voice quality, latency, and generation reliability.

We have deprecated the multiNativeLocale field across all Text-to-Speech (TTS) endpoints, including:

• Non-Streaming API
• Streaming API
• WebSockets

The multiNativeLocale field has been replaced by the locale field. We recommend updating your integrations to use the locale field to ensure continued compatibility with future updates.

A new optional model query parameter has been added to the /v1/speech/voices endpoint.

By default, this endpoint returns all available GEN2 voices. When you specify the model parameter, the response returns voices appropriate for the specified TTS model. Valid values are FALCON and GEN2.

• Sub-130ms time-to-first-audio for instant responses
• Multilingual speech - Handles mixed-language sentences effortlessly, with native-like pronunciation and prosody.
• Expressive prosody with 99.37% pronunciation accuracy
• Data residency in 11 regions via streaming & WebSocket TTS endpoints

Learn more about Falcon here.

Murf API now supports 14 additional languages, expanding its global reach and accessibility. The newly added languages are:

• Malay
• Filipino
• Czech
• Finnish
• Thai
• Vietnamese
• French Canadian
• Swedish
• Telugu
• Malayalam
• Kannada
• Marathi
• Gujarati
• Punjabi

These languages come with high-quality voices and support for advanced features like Multilingual technology and voice customization. Explore the new voices in our Interactive Voice Explorer or in our Voice Library.

We’ve enhanced data privacy by allowing zero data retention. By setting the encodeAsBase64 parameter to true, you can receive synthesized audio as a Base64 encoded string directly in the API response, with zero data stored on Murf’s servers.

This feature is designed for applications handling sensitive information, ensuring that your data is not persisted after the API call.

A new parameter, wordDurationsAsOriginalText, has been added to the /v1/speech/generate endpoint.

By default, the wordDurations object in the API response contains normalized text. When you set wordDurationsAsOriginalText to true, the response will instead include the original, un-normalized text from your request. This allows for a direct mapping between your input text and the corresponding word-level timestamps.

We’ve enhanced our audio format support for both Text-to-Speech and Streaming APIs to provide you with greater flexibility.

• The format parameter in the Text-to-Speech API (/v1/speech/generate) now accepts PCM and OGG.
• The format parameter in the streaming API (/v1/speech/stream) now accepts PCM.

These new options are available for immediate use.

We are excited to launch WebSocket streaming for our Text-to-Speech API. This feature enables developers to build real-time, low-latency voice experiences by streaming text and receiving synthesized audio over a persistent, bidirectional connection. It is ideal for applications like conversational AI, live chat support, and dynamic content narration where immediate audio feedback is crucial.