ASR is the process of converting spoken audio into text using AI models. In VoIP telephony, ASR sits downstream of audio capture – typically SIPREC-based media forking from a PBX or SBC – and converts RTP audio streams into transcripts.

Telephony ASR is harder than transcribing a podcast: the audio is often narrowband at 8kHz, subject to packet loss, may carry multiple speakers, and includes background noise and codec artifacts. ECG accounts for these obstacles and builds AI ASR systems that actually work in real-world telephony environments.

ASR and AI automated speech recognition are essentially the same thing. The term AI ASR or AI automated speech recognition emphasizes that the system uses artificial intelligence and deep learning models rather than older rule-based or template-matching approaches.

We use the Hugging Face Open ASR Leaderboard to identify the leading open source options. As of 2026, we use NVIDIA Parakeet for English telephony efficiency, Whisper Large V3 Turbo for multilingual support across 100 languages, IBM Granite Speech 3.3 for high-accuracy regulated industries, and Moonshine for low-latency streaming.

Choosing the right ASR model will depend on your workload. We assess your needs and evaluate against your actual audio before committing to an architecture. 

vCon (Virtual Conversation) is an emerging IETF Internet standard for encapsulating conversation artifacts, including audio recordings, transcripts, metadata, and analysis, in a structured, interoperable format.

vCon enables interoperability between different call recording and AI processing platforms. We implement vCon as the internal data structure so your platform can ingest recordings from third-party compliance recording systems without custom integration for each vendor.

Accuracy depends heavily on audio quality. Noisy contact center audio, strong accents, or industry-specific terminology can drop accuracy, which is why it's important to evaluate ASR models against your actual call data before deployment.

Yes. The primary method for accomplishing this is SIPREC (SIP Recording), which is an IETF standard protocol that forks a copy of RTP audio from a call to a recording server.

ECG's platform acts as the SIPREC endpoint. It accepts the SIPREC INVITE from the SBC, negotiates media, and receives a one-way audio stream. The media fork happens at the SBC before reaching any other system, so call quality for the actual participants is unaffected regardless of what happens in the transcription pipeline.

Vision and speech processing in artificial intelligence refers to multimodal AI systems that combine computer vision (understanding images and video) with speech processing (understanding audio).

Together, vision and speech processing enable applications like video conferencing with automatic captions, or analyzing recorded videos with both visual and audio understanding.

Both image and speech processing rely on deep learning models trained on large datasets. For speech, the models learn to recognize patterns in audio waveforms and convert them to text. For images, models learn to recognize visual patterns and extract meaning. What enables them is access to training data, computing power for training and inference, and open source frameworks like PyTorch and TensorFlow.

In voice applications, speech processing is the focus. Multimodal systems that combine vision and speech exist, but telephony solutions typically only require speech understanding.

There are four big ones:

1. Benchmark accuracy doesn't match real-world performance. Models tested on clean audio perform worse on telephony's narrowband codec and background noise.
2. Production deployment takes about 3x the effort of initial prototyping once you add streaming, VAD, diarization, and system integrations.
3. Most SaaS transcription routes audio to third-party clouds, which breaks HIPAA, GDPR, and data governance requirements. You need an on-premises deployment for regulated data.
4. SIPREC vendor implementations vary from the standard, so you have to normalize XML handling across different SBCs.