Sponsor of the year's leading signal-processing conference

Olewave was a patron and exhibitor at IEEE ICASSP 2024 in Seoul, alongside Meta, Google, Samsung, Apple, ByteDance, Adobe, MathWorks and many others. See the full patron list →

Thanks to everyone who stopped by our booth to talk speech, data, and voice AI.

Spotlight talk by our Founder & CEO

Wei Chu, Founder & CEO of Olewave, delivered a spotlight talk during the conference:

Harnessing Wild and Untamed Data for the Cost-Efficient Development of Customized Speech-Related Services

Abstract. We recently discovered that models trained with large-scale speech datasets sourced from the web could achieve superior accuracy and potentially lower cost than traditionally human-labeled or simulated speech datasets. We developed a customizable AI-driven data-labeling system. It infers word-level transcriptions with confidence scores, enabling supervised ASR training. It also robustly generates phone-level timestamps even in the presence of transcription or recognition errors, facilitating the training of TTS models. Moreover, it automatically assigns labels such as scenario, accent, language, and topic tags to the data, enabling the selection of task-specific data for training a model tailored to that particular task.

We assessed the effectiveness of the datasets by fine-tuning open-source large speech models such as Whisper and SeamlessM4T and analyzing the resulting metrics. In addition to openly-available data, our data-handling system can be tailored to provide reliable labels for proprietary data from certain vertical domains — enabling supervised training of domain-specific models without human labelers, eliminating data-breach risk, and significantly reducing labeling cost.

View talk details on the ICASSP 2024 site →

What we showcased

Speech datasets

• Large-scale — up to millions of hours
• Real-world scenarios — meetings, calls, talks, natural conversations
• Validated verbatim transcriptions with word- and utterance-level confidence
• Multi-language and multi-region — actively collecting Brazilian Portuguese, Latin American Spanish, Arabic, Southeast Asian languages, Chinese, Japanese, and Korean
• Diverse topics — fashion, entertainment, healthcare, and more

Beyond ASR, we also produce corpora for meeting summarization, speech synthesis and voice cloning, and evaluation-only sets for benchmarking your own models. We deliver high-fidelity audio/video for realistic speech and talking-head synthesis.

Bespoke solutions for proprietary data labeling

Our four-step pipeline lets you label sensitive data without ever exposing it to a human labeler:

1. Jump Start — fine-tune a domain model on wild data from a domain matching yours, starting from a large speech/language foundation model.
2. Auto Label — the domain model labels your proprietary data with confidence scores. No human labelers. No data-breach risk.
3. Iteratively Learn — combine labeled private and public data to sharpen the model, then re-label your private data with the better version.
4. Good to Go — when quality plateaus, we license our Tycho SDK so you can keep labeling and building on your own infrastructure.

Tycho — our SDK for in-house speech services

Tycho is the SDK we exhibited at ICASSP for teams building high-ROI, in-house speech products.

• Always state-of-the-art in model accuracy and inference speed
• Support for training, fine-tuning, evaluation, and deployment
• Available models: online / offline ASR, speech assessment, voice cloning — new models available on request
• Runs on modest GPU workstations to keep cost low
• All training, fine-tuning, evaluation, and deployment code is delivered to the client

Competitive service rates for:

• Fine-tuning your vertical-domain model to top-notch quality
• Fixing and improving a launched service via deep bad-case analysis
• Hands-on training for your ML engineers on speech R&D projects

Tycho is not open-source. Ask us for a licensing quote and service charge.

Come find us next time

If you missed us in Seoul, we'll be at more conferences later this year. Email us at [email protected] or reach out via the website — datasets, custom models, or consulting, we'll respond within one business day.