Finch-Global
AI model · Speech to text · Audio / speech → Text by Murf
Real-time transcription for global voice agents and telephony.
A real-time speech-to-text model for English and European languages, designed for voice agents and telephony with word-level timestamps and low-latency streaming.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
- Price
- From 0.12 USD per hour
- Languages
- English (US), English (UK), English (Australia), German, Spanish, French, Italian, Portuguese, Dutch
Category: Conversational AI
- Modalities
- Audio (Live Stream) → Text
- Real-time streaming
- Real-time streaming over WebSocket
- Access
- API (WebSocket)
- Fine-tuning
- Keyword boosting for brand names and domain vocabulary
- Model ID
- finch_global_1.0
- Host URL
- us-east.api.murf.ai
- Real-time streaming via WebSocket
- Optimized for English and European languages
- Telephony-ready 8 kHz µ-law and A-law support
- Word-level timestamps on final transcripts
Overview
Finch-Global is a real-time speech-to-text model developed by Murf AI, specifically optimized for English and European languages. It is designed for businesses building voice agents, IVR systems, and call analytics tools that require low-latency transcription. The model is hosted in the US East region and supports multiple English dialects alongside major languages such as German, Spanish, and French. It handles telephony-grade audio directly, including 8 kHz µ-law and A-law formats, and provides word-level timestamps for precise search and review. Finch-Global operates over a single WebSocket connection, delivering partial transcripts in real time and final transcripts when the speaker pauses.
Transcription Capabilities
- Real-time Streaming: The model converts live audio into text in real time over a single WebSocket connection.
- Multi-turn Sessions: The server ends each turn when the caller pauses, maintaining a single session timeline for multi-turn calls.
- Partial and Final Transcripts: Users receive partial transcripts while the caller is speaking and a final transcript when the turn ends.
Languages and Detection
- Global Language Support Office: The model is optimized for English (US, UK, Australia) and major European languages.
- Automatic Language Detection: The system can detect the language from the audio automatically among its supported languages.
- European Language Coverage: Supported languages include German, Spanish, French, Italian, Portuguese, and Dutch.
Technical Features
- Telephony-ready Audio: The model accepts 8 kHz µ-law and A-law telephony audio directly from providers.
- Background Voice Cancellation: Users can set a background voice cancellation parameter to remove other voices before transcription.
- Keyword Boosting: The model supports keyword boosting to improve recognition of brand names and product terms per connection.
- Flexible Audio Formats: Audio can be sent as binary frames in PCM and Opus formats at rates up to 48 kHz.
Use cases
- Developers build voice agents that show a caller's words as they speak and pass completed turns to an LLM.
- Financial institutions use keyword boosting to improve recognition of product names in payment reminder calls.
- Customer support teams transcribe calls in the caller's native language using automatic language detection.
- Quality assurance teams use word-level timestamps to search and redact specific terms in call recordings.
Specifications
- Modalities
- Audio (Live Stream) → Text
- Real-time streaming
- Real-time streaming over WebSocket
- Audio formats
- 8 kHz µ-law, 8 kHz A-law, 16-bit PCM (up to 48 kHz), Opus (up to 48 kHz)
- Access
- API (WebSocket)
- Fine-tuning
- Keyword boosting for brand names and domain vocabulary
- Languages
- English (US, UK, Australia), German, Spanish, French, Italian, Portuguese, Dutch
- Model ID
- finch_global_1.0
- Host URL
- us-east.api.murf.ai
- Hosting Region
- US East
- Pricing Unit
- $0.12 per hour (billed per second)
- Default Audio Format
- 16-bit PCM, 16 kHz, mono
- Max Sample Rate
- 48 kHz
- Telephony Formats
- 8 kHz µ-law, 8 kHz A-law
- Connection Type
- WebSocket (WSS)
How it works
- Connect to WebSocket: Open a WebSocket connection to the US East host and pass the model name along with your API key.
- Stream Audio Frames: Send audio as binary frames; the server returns partial transcripts as the turn grows.
- Receive Transcripts: The server detects pauses to end turns and provides final transcripts with word-level timestamps.
Pricing
- Speech to Text Beta Usage
- $0.12 per hour
FAQ
What is the pricing for Finch-Global?
Speech to Text costs $0.12 per hour of audio, which is billed per second of usage.
Which languages does Finch-Global support?
Finch-Global is optimized for English (US, UK, Australia) and European languages including German, Spanish, French, Italian, Portuguese, and Dutch.
What audio formats are supported for telephony?
The model accepts 8 kHz µ-law or A-law telephony audio directly, as well as PCM and Opus formats at rates up to 48 kHz.
Does it support automatic language detection?
Yes, the model can detect the language from the audio automatically if the language parameter is omitted or set to auto.
Added by AntSpace from public sources (murf.ai).