
Voice AI Weekly #17 - Third Week of November 2026
In the global voice AI industry this week, there was a noticeable push toward improving STT accuracy while expanding deployment options, such as on-premises and on-device. Here is a summary of the five key stories highlighted by MAGO.
1. ElevenLabs Unveils Scribe v3
On November 10, ElevenLabs unveiled its speech-to-text model Scribe v3 along with its real-time version, Scribe v3 Live. Across 102 languages, it reduces the word error rate (WER) by 27% compared to the previous version, while maintaining the same pricing. It seamlessly handles code-switching utterances where the language changes mid-sentence. This marks an instance where a company known primarily for text-to-speech (TTS) is expanding its footprint into the STT domain.
2. Salesforce Releases Open-Weight STT SOV-Transcribe-1
On November 12, Salesforce released the open-weight speech recognition model SOV-Transcribe-1 under the Apache 2.0 license. Combining a Whisper large-v3 encoder with the Gemma 4 LLM, it supports seven languages and achieved a 2.28% WER according to Artificial Analysis benchmarks, taking first place among open-weight models.
3. Pipecat 1.16 Adds Audio Classifier Detecting Tone, Laughter, and Sarcasm
The open-source voice agent framework Pipecat released version 1.16.0 on November 12. It includes an alpha feature based on the open model Pipecat Voice Reader that classifies a speaker's tone, laughter, sarcasm, and speaking rate. The classification results, along with output from Hume's Audio Decisions API, are passed to the LLM on every turn. With text transcripts alone, it is difficult to determine whether a phrase like "Yeah, great" is genuine positivity or sarcasm. Efforts to directly extract nonverbal cues from voice signals and incorporate them into agent responses are gaining real momentum.
4. Google Adds Offline Speech Recognition and TTS to AI Edge Gallery
On November 10, Google updated its AI Edge Gallery app with on-device voice capabilities. Users can run speech recognition in over 100 languages and TTS with eight voices across 21 locales without a network connection. It leverages ML Kit's GenAI Voice API on Android and open models such as Parakeet-TDT 0.6B and Kokoro-82M on other platforms. Major platforms have begun offering on-device voice stacks to developers out of the box.
5. Microsoft Officially Launches Azure Real-Time Multichannel STT
Microsoft has officially launched real-time multichannel speech recognition for Azure Speech. It can recognize up to two separate channels in a single session, allowing agent and customer voices to be processed separately. Different languages can be designated for each channel, with the pricing remaining identical to single-channel usage. The starting point for contact center call analysis lies in accurately distinguishing speakers. Clear channel separation is essential to accurately derive emotion or stress metrics specifically from the customer's voice.
If you are considering contact center voice analytics or adopting on-premises or on-device voice AI, please feel free to reach out to us at contact@holamago.com. We will be back next week with more global voice AI news curated by MAGO.
