InteractionBench measures timing, task accuracy, and whisper in real-time voice systems using 221 items across thirteen scenarios. The benchmark evaluates six full-duplex systems and reports a timing-task accuracy rank correlation of 0.83.
Key points
221 items per system in thirteen scenarios.
Timing and spoken-task accuracy rank-correlated (ρ = 0.83).
Sources
Zero-Shot Voice Cloning cuts children's ASR WER to 16.4%
Zero-Shot Voice Cloning for Synthetic Data Augmentation in Children's ASR reduces word error rate to 16.4% on the ENNI corpus. The study uses OmniVoice TTS to generate up to 16,000 child-like utterances and fine-tunes Whisper-small under various data conditions.
Key points
Synthetic data alone lowers WER from 19.8% to 16.4% on ENNI.
Synthetic + real data fine-tuning yields 9.2% WER, matching real-data-only fine-tuning at 8.8%.
Sources
SoundHound AI and LG Partner to Expand Commerce on 20M Smart TVs
SoundHound AI and LG announced a partnership to launch a commerce channel on about 20 million LG Smart TVs in the U.S., starting with restaurants. The program will let participating brands reach households directly through TV screens without new hardware.
Key points
Initial footprint: approximately 20 million LG Smart TVs in the United States.
SoundHound currently works with over 15,000 restaurant locations across the U.S.
Sources
Soniox STT v6 launches with agent text input and echo cancellation
Soniox STT v6 is released with agent text input that provides recognition context from both sides of a conversation and agent voice echo cancellation that filters out the agent’s playback while listening to the user. The model supports real-time streaming and asynchronous transcription across 60+ languages and adds stronger speaker diarization, language identification, endpoint detection, and context customization.
Key points
Agent text input streams the agent’s outgoing text into the model to improve recognition of user responses.
Agent voice echo cancellation removes the agent’s playback from recognition while still listening for user interruptions.
Sources
- Original post
- Soniox Introducing Soniox STT v6 - Linked page
Gradium HDS certification enables health data hosting
Gradium is now HDS certified, allowing healthcare teams to build patient calls, appointment booking, medical transcription, and more. The certification is France’s standard for hosting health data.
Key points
Certification: Gradium is now HDS certified.
Capabilities: Healthcare teams can build patient calls, appointment booking, medical transcription, and more.
Sources
- Original post
- Gradium Trust Center - Linked page
How Much Prosody Does Your ASR Encoder Already Know
How Much Prosody Does Your ASR Encoder Already Know shows prosody adds a large gain on raw audio but the gain collapses after the first encoder layer. The paper tests three ways on Zipformer, wav2vec2 and Whisper encoders and finds the same pattern across them.
Key points
Prosody adds a large gain on raw audio.
The gain collapses once the signal passes through even the first encoder layer.
Sources
Audio Data for AI Training – Refinery launch announced
Audio Data for AI Training introduces the Refinery, a service that converts mixed-audio files into usable data for labs, archives and data marketplaces. It claims to unlock real conversations from a single mixed recording.
Key points
The Refinery unlocks labs, archives, data marketplaces from the audio you already have.
It works on real conversations mixed into one file.
Sources
- Original post
- Audio Data for AI Training - Linked page
Count-Density Networks for Modal Plate Estimation
Count-Density Networks for Modal Plate Estimation described two neural methods for Task B of the 1st DAFx Parameter Estimation Challenge. The real-valued convolutional network uses pooled spectral features and absolute-scale conditioning; the complex-valued Transformer network processes complex inputs. Both jointly infer modal count and per-mode attributes from the impulse response and achieve lower overall challenge error than classical baselines, with frequency and decay estimates substantially more accurate than gain estimates.
Key points
Real-valued convolutional count-density network combines pooled spectral features with time-domain and absolute-scale conditioning.
Complex-valued Transformer count-density network processes complex inputs and jointly infers modal count and per-mode attributes.