LLM-based speech systems are fluent, and they guess. Ask one for a part number, a customer's surname or a serial code and it returns something plausible. Inferret recognises your vocabulary: the fixed lists, proper nouns and alphanumeric codes your business actually runs on. We complement the large models rather than compete with them, and we build the training data behind both.
Trusted by
LINE Yahoo
Hitachi Systems
UsedCar.com
JIEM
Best Path Research
Large models transcribe conversational speech beautifully. They are far less reliable the moment a word has to be exactly right rather than merely plausible, because they predict the likeliest next token instead of matching against a known set. Feed SpeechFoundry your vocabulary and it recognises against that list, so the output is constrained to values your systems can actually accept.
Supply the words that matter: part numbers, SKUs, serial ranges, customer surnames, drug names, station names, model codes. Recognition is constrained to that set, so results land on a real record instead of a convincing near-miss.
Strings like "K7-942A" or a fifteen-digit serial defeat general transcription, which renders them phonetically and inconsistently. We handle spoken letters, digits and separators as structured values, and validate them against the format you define.
A user says a fragment of a long proper noun and still resolves to the right entry. Useful wherever names are long, formal or awkward to say in full: phonebooks, track titles, restaurant and business names.
Because the vocabulary is defined rather than inferred, behaviour is testable and repeatable. You can prove what the system will and will not accept, which matters in regulated and safety-critical work where an LLM's confident invention is unacceptable.
In practice the two approaches sit side by side: a large model handles open conversation and reasoning, while SpeechFoundry handles the fields that must be exact. We routinely build systems that route between them.
Take the whole platform or the one component you're missing. Every piece is customised to your domain. That is the part that decides whether voice actually works.
Give a device or an application a spoken interface. Users say what they mean, like "it's too hot in here", and the system resolves intent, holds multi-turn context and acts. Language-model reasoning where it helps; deterministic grammars where you need guarantees.
Spoken queries against large catalogues and proper-noun databases: places, products, songs, people. Fuzzy and partial matching means a user can say part of a long name and still land on the right record.
High-accuracy transcription of large volumes of audio and video, in real time or batch, with optional human verification. Search the transcript, jump straight to that moment in the audio, and get alerted the instant a keyword or phrase is spoken.
Reasoning, summarisation, extraction and dialog on top of recognition, using open models you can host yourself where privacy or cost demands it. We fine-tune on data built for your domain, keep the deterministic parts deterministic, and evaluate against a benchmark rather than a demo.
Human-collected and synthetically generated datasets, annotated to spec, for speech and language models, ours or yours. See how we build it →
Problems have voices too. The same acoustic modelling that recognises speech can learn what a healthy machine sounds like and flag the moment it stops: a wind-turbine rotor changing pitch, a loose part on an assembly line, an unusual noise outside a building.
Every model we have built in nearly two decades ran into the same bottleneck. Not architecture: data. So we build data as a product in its own right, collected from real people, generated synthetically at scale, or blended, then annotated and quality controlled to a standard you can actually train on.
Native speakers and domain specialists producing the ground truth a model is measured against. Speech collection in the acoustic conditions your product actually meets, transcription to a defined style guide, intent and entity labelling, preference and ranking judgements for model alignment, and red-team prompts written by people trying to break things.
Language models generating corpora under constraint, speech synthesised across voices and accents, and acoustic augmentation that puts clean audio into noisy rooms, moving cars and telephone channels. The point isn't volume for its own sake. It is deliberate coverage of the long tail and the edge cases your real data will never contain enough of.
Schema, label taxonomy and quality bar agreed before anyone annotates anything.
Human collection, synthetic generation, or a deliberate mix of the two.
Labelled against the guidelines, with agreement measured rather than assumed.
Held-out evaluation sets and benchmarks, so you can prove the data moved the metric.
Your format, your infrastructure, with provenance and licensing documented.
Instruction, preference and evaluation data for adapting a language model to a domain, including the awkward cases where a general-purpose model quietly gets things wrong.
Held-out test sets built to expose real failure modes rather than flatter a model. If you can't measure it honestly, you can't improve it.
Provenance tracked and consent documented, so what you train on is defensible, and increasingly the part that decides whether a model can ship at all.
Designed for low signal-to-noise conditions such as moving vehicles, restaurants and station concourses, not quiet demo rooms.
People speak the way they'd speak to a person. No memorised command list, no unnatural phrasing.
Language models of a million distinct words and beyond, in both embedded and connected deployments.
Always-listening activation on a keyword of your choosing, removing the press-to-talk button entirely.
Users interrupt a prompt mid-sentence and are heard immediately, the way a real conversation works.
Models tuned for telephony as well as PC and mobile audio, so call-centre channels are not an afterthought.
Multi-step conversations that retain what was said earlier, so follow-up questions don't start from zero.
Models tune to a primary user's voice and pronunciation over time, and accuracy climbs with use.
Voices are as distinctive as fingerprints, so you can tell your users apart and personalise accordingly.
Accuracy, speed and memory traded off deliberately for your target hardware rather than left to chance.
Voice and language are what make us human. We build the stack that lets a machine take part in that properly: hearing accurately, grasping meaning, and answering in kind.
Deep neural network acoustic models combined with our proprietary, patented weighted finite-state transducer compression. That is the reason we get high accuracy and fast response inside a memory budget that would normally rule both out.
Recognition turns sound into text. It does not tell a machine what the words mean. Our understanding layer closes that gap: users say things as they come to mind, and the system resolves the intent, the entities and what to do next, without forcing anyone to memorise a command list.
Where a task calls for open-ended reasoning or generation, we bring language models into the pipeline, including open models hosted on your own infrastructure, while keeping recognition and intent handling deterministic and testable. Fine-tuned on data built for your domain, and measured against a benchmark we build alongside it.
Every component is built for your project. Generic "cover-everything" APIs cannot be tuned to your vocabulary; a model built around your domain terms, your product names and your users' phrasing can, and that is where the accuracy gap opens up.
Text-to-speech, dialog management, context management and named entity recognition over very large proper-noun lists: the components needed to close the loop and let a device hold its side of a conversation.
Because the full engine runs on-device, privacy isn't a policy promise. It is an architecture. Regulated, safety-critical and offline products get the same capability as connected ones, with audio that never leaves the box.
All major world languages are supported in production, with ready-made generic language models and large vocabularies. Every model comes in wideband for PC and mobile, and narrowband for telephony.
Including
Need one we don't already run? We have a clear, transparent process for building a production-quality acoustic model for a new language in two to three months, depending on the availability of suitable training data, plus mechanisms to bootstrap a prototype far sooner for testing. Where the data doesn't exist, we can create it: see training data and annotation above. Tell us which languages you need and when, and we'll come back with a plan.
Most voice stacks force a choice: a small, limited recogniser on the device, or a large capable one in someone else's cloud. SpeechFoundry runs identically in both, with the same core, the same acoustic models and the same language models. You are not locked into where your audio is processed, and you can change your mind late in the product cycle without rewriting anything.
DNN acoustic models · patented WFST compression · constrained vocabulary decoding · natural language understanding · dialog & context
Ships as an SDK for your platform. Talks to your software over a local API and returns plain JSON. Nothing goes to a network.
The same engine as a service, on dedicated servers in your infrastructure or ours. Audio and JSON results over standard protocols.
Understanding voice makes work easier almost everywhere. Our team has put voice into production across a wide span of environments, most of them noisy, constrained, or both.
Embedded voice with wide domain coverage and high accuracy inside the car, where other vendors fall back to a cloud connection they can't guarantee.
Noise-robust, multilingual guidance on connections and nearby points of interest for travellers in busy metro stations.
Real-time, high-accuracy subtitles for news and live programming.
Automated agents that handle routine calls end to end, plus intent analysis across very large volumes of recorded conversation.
Voice replacing fiddly touch interaction in iOS and Android products.
Sessions, speeches and panels transcribed live into searchable, formatted text, with instant alerts when a chosen keyword is spoken.
Language models customised to a clinical domain's terminology so reports can be dictated and searched automatically.
Damage descriptions and assessments captured by voice dialog on-site, so an assessor never has to look away to fill in a form.
Pronunciation recognised and scored, so a learner gets feedback on how they actually sound.
Lights, climate and appliances, all controlled by speech, embedded or connected, with the same engine either way.
Machinery, rotors and processes monitored acoustically, with alerts the moment the sound signature drifts.
High-volume batch recognition over recorded interactions, processed inside your own infrastructure.
Not every customer needs the same thing from voice, and nobody gets there by having a technology thrown at them. Customisation and support are part of every engagement, whether you have in-house speech experts or this is your first time near the subject.
You have a product driven by screens and buttons and you want to add voice, or replace that interface with it. We gather the requirements, suggest what voice should and shouldn't control, and build a prototype that shows you the difference. If it works for you, we take it from prototype to a production-grade voice interface.
The single biggest lever on real-world accuracy. We build language and semantic models around your specific domain rather than handing you a generic model and hoping. This is exactly what the free voice APIs cannot do.
Create, customise and test voice solutions in our browser-based development environment, free, with nothing to install. Get an account, do some initial setup with our team, then start building voice commands and testing them immediately with your own microphone. When you're ready, we help you optimise and port what you built to your embedded or production environment.
Upload audio or video and get transcripts back. Search them, jump to the relevant moment in the recording, or run the whole pipeline on your own hardware so that nothing ever leaves your building. That last option matters for legal, clinical and HR recordings, where sending audio to a third party is often simply not permitted. We can fold your domain vocabulary into the dictation models and set up keyword alerts that flag or email you the moment a phrase is spoken.
A dataset built to your specification: human collection, synthetic generation, annotation, QA and a held-out evaluation set, whether the model it feeds is ours or entirely your own. This is the fastest-growing part of what we do, and often the highest-leverage thing a team can outsource.
We also improve voice built on someone else's stack. Our team has customised and deployed solutions across a range of third-party voice platforms and assistant frameworks. If your product already has voice and the user feedback is poor, that is a solvable problem and often a fast one.
Inferret is a small, agile, international group of researchers and engineers who have been building speech and language technology since 2007, through the statistical era, the deep learning shift, and now the language model era. We own our engine: the acoustic models, the decoder, the compression method, the understanding layer.
That matters more than it used to. When the interesting part of a product is voice or language, depending on an API you cannot inspect, tune, or run on your own hardware is a structural weakness. We build the parts you need to control, and we make them yours.
Having trained models across many languages, we know where they break, and it is almost always the data. That is why dataset construction and annotation now sit alongside the engine as a business in their own right, for teams training their own models as much as for teams using ours.
Work with usWhether you know exactly which component you need or you're still working out whether voice belongs in your product at all, start here. We'll answer with something useful, not a brochure.
