On-prem speech AI that hears Taiwan, running in your own server room — Mandarin, Taigi, Hakka, Indigenous languages and English
S1彼个案子的預算,咱敢有確定落來?Has the budget for that project been finalized?
S2上禮拜董事會通過了,我等一下把會議紀錄寄給你。The board approved it last week — I'll send you the minutes shortly.
S1好,數字若有 update 你直接 tag 我就好。OK — if the numbers change, just tag me.
S2沒問題,摘要跟待辦我一起附上去。No problem, I'll attach the summary and action items too.
Mandarin, Taigi, Hakka, Indigenous languages and English — all transcribed into Traditional Chinese, with every model running on your own servers.
01Data boundary
Audio never uploaded.
Transcripts never leave your server room.
- MODEL 1Speech recognition
- MODEL 2Diarization
- MODEL 3Dialogue
- MODEL 4Speech synthesis
- Audio
- Transcripts
- Summaries
- Voiceprints
- Domain dictionaries
Not a single clip of audio, transcript or prompt goes out.
02Products
Three products, one principle.
MosaTalk
On-prem AI voice agentAbout one second from when the caller stops talking to when the AI answers — and callers can interrupt anytime.
MosaMinutes
On-prem AI meeting minutesTranscript, speakers, summary and action items — edit while you listen until it's ready to send.
MosaScore
On-prem call analyticsEvery call heard in full, every utterance tagged with emotion and intent.
03Numbers we've run ourselves
From the end of speech to the AI's reply
Live transcription on one NVIDIA DGX Spark GB10
Built in, all transcribed into Traditional Chinese
Runs the whole system, recognition and summaries included
04Reproducible evaluation · coming soon
"What's your recognition accuracy?"
CER = %
Marketing numbers mean nothing. Reproducible numbers do.
The number that matters most is the one you measure on your own recordings in a proof of concept.
Product 01 · On-Premise
MosaTalkAn on-prem AI voice agent that understands Taigi
Customers tap Call on a web page or dial your existing support line, then simply talk to the AI.
- ListenFires the moment you finish; waits a beat when you hesitate — never cuts in
- ThinkAnswers directly, only from the material you provide
- SpeakStreams sentence by sentence, so the first one goes out fast
- 01Barge-in anytimeThe moment a caller speaks, the AI stops and listens.
- 02No code to update knowledgeLoad a simple Q&A document and you're done.
- 03Never makes things upIf the answer isn't in your material, it takes a callback number instead.
- 04Your brand's voiceClone a custom voice from a single reference recording.
- Languages
- Mandarin, Taigi and English — understands Taigi, replies in Mandarin for now
- Access
- Browser calls with nothing to install, or your existing phone line
- Hardware
- One NVIDIA GPU, validated on a single 46 GB card
In developmentTaigi speech synthesisHakka supportHuman handoffBusiness-system API callsCall analytics dashboard
SpecsSpecs & limitations
- Languages
- Mandarin, Taigi and English, with automatic language detection. For now it understands Taigi but replies in Mandarin; Taigi synthesis and Hakka are still in development, and Indigenous languages aren't in the voice agent yet.
- Access
- Standard WebRTC for browser calls with nothing to install, or your existing support line through a telephony provider — a phone integration is already deployed in production with a telephony partner.
- Answer policy
- Answers only from the material you provide, with numbers stated exactly; short and conversational, three sentences at most. If something isn't covered, it never makes it up and asks for a callback number instead.
- Outbound calls
- Besides answering calls, it can place them: the AI opens the conversation as soon as the call connects — useful for appointment reminders.
- Deployment modes
- Standalone: one GPU server is the whole system, intranet only. Edge-split: the GPU host stays inside your network with no fixed IP and no open ports, but it also needs one small cloud host.
- Hardware
- Ubuntu Linux x86_64 with one NVIDIA GPU (validated on a single 46 GB card); recognition, dialogue and synthesis share the card through planned memory allocation.
- Observability
- An eight-color live event stream, so acceptance testing and troubleshooting aren't a black box.
- Open standards
- WebRTC and an OpenAI-compatible interface; models can be swapped or upgraded, and sourced to match procurement rules.
FAQ
How is an AI voice agent different from a push-button IVR?
An IVR makes callers press 1 or 2 through a menu, and one wrong turn means starting over. MosaTalk listens to what the caller says and answers in natural language — about one second after they stop talking — and callers can interrupt at any time.
Can the AI answer our phones, including our existing support line?
Yes. Customers can call from a web page in any modern browser that supports WebRTC, with nothing to install. Through a telephony provider integration, customers can also dial your existing support number from a landline or mobile and the AI answers.
Does it understand Taigi? Can it reply in Taigi?
It understands Taigi, including natural Taigi–Mandarin code-switching — the model is optimized for Taiwanese accents and Taigi. Taigi speech synthesis isn't available yet, so for now it understands Taigi and replies in Mandarin. Taigi synthesis and Hakka support are still in development, not deliverable features.
Are call recordings uploaded to the cloud?
No. Recognition, dialogue and speech synthesis all run on your own servers, with no calls to any cloud AI service. Standalone mode needs only your intranet; edge-split mode also needs one small cloud host.
What hardware do we need?
Ubuntu Linux x86_64 with one NVIDIA GPU with plenty of memory; it's validated on a single 46 GB card. We can supply the hardware and install it on site.
Does updating the knowledge base require coding?
No — a Q&A-format document is enough. When a question isn't covered by your material, the AI doesn't make up an answer; it asks the caller to leave their contact details.
Test it with your own knowledge base and real customer calls.
Request a PoC
Product 02 · On-Premise
MosaMinutesOn-prem AI meeting minutes
Record or upload, and get a transcript, who said what, a summary and action items — without anything leaving your network.
彼个案子的預算,咱敢有確定落來?
上禮拜董事會通過了,我等一下把會議紀錄寄給你
好,數字若有 update 你直接 tag 我就好。
ACTION ITEMS
- 李專員寄送董事會決議之會議紀錄
- 預算數字更新後通知王主任
- 01Five languages, one transcriptMandarin, Taigi, Hakka, Indigenous languages and English, all into Traditional Chinese.
- 02Remembers who's whoEnroll a voiceprint once; the same speaker is recognized from then on.
- 03Edits never get overwrittenFour independent editing channels that never clobber each other.
- 04Ready to sendExport to Word, plain text or subtitles, with action items as checklists.
- Capacity
- 10 live streams on one NVIDIA DGX Spark GB10
- Deployment
- Containerized, runs on a single machine, works air-gapped
- Admin
- OIDC single sign-on, tiered permissions, audit logs
In developmentAutomatic deliverySensitive-data maskingDirect online-meeting captureGovernment document formatCross-meeting Q&A
Field-testedProblems fixed in production
The scariest failure isn't mishearing — it's inventing words.
All three of these were found and fixed while processing real meetings in production, not written into a spec sheet.
- 01Short noises aren't treated as speechA minimum speech length keeps 200 ms of background noise out of recognition — and off the bill.
- 02No repetition loopsWe found the root cause and fixed it at the system level, rather than hiding it with after-the-fact filtering.
- 03Recognizer confidence is loggedAbnormal output is traceable and auditable, never silently mixed into the transcript.
SpecsSpecs & limitations
- Capacity
- One NVIDIA DGX Spark GB10 handles 10 live streams; uploads use a batch queue with no stream limit; scales out to multiple machines beyond that.
- Online meetings
- It can't connect directly to Google Meet, Microsoft Teams or Zoom yet — record or upload the audio instead. Direct capture of online-meeting audio is on the roadmap.
- Languages
- Mandarin, Taigi, Hakka (six accents), Indigenous languages (42 across 16 peoples) and English, all into Traditional Chinese with automatic punctuation; users can set a default language and accent.
- Speakers
- Voiceprint enrollment and automatic matching; you can set the expected number of speakers on upload; if speaker ID fails, you still get the full transcript.
- Export
- Word (with attendees, a summary section and action-item checklists), plain text and SRT subtitles.
- Accounts & audit
- OIDC single sign-on (self-hosted identity providers supported), tiered permissions, admin actions logged with a required reason, and configurable result retention.
- API
- Access tokens for internal systems, synchronous and asynchronous; duplicate requests run only once. Usage is metered separately: quotas for browser users, credits for API calls.
- Data integrity
- Every recording is validated before it enters the system; its hash and byte count are logged on receipt, so the original audio's integrity is verifiable.
- Deployment
- Containerized, runs on a single machine, works air-gapped. Users need only a modern browser, on desktop or mobile.
- Model sourcing
- Built on industry-standard open-source technology and compatible interfaces; where procurement rules restrict model origin, models can be swapped for open-weight models from other sources.
More than software
- 01Hardware and on-site installationTested and installed for you.
- 02Continuous model improvementTuned to your real audio conditions and terminology.
- 03Reproducible evaluation reportsOn public benchmarks you choose or your own meeting audio.
- 04Compliance and supply-chain fitAccess control and audit tailored to your rules.
- 05Scenario setupYour minutes format and glossary turned into configuration.
- 06Maintenance, updates and trainingNo in-house AI/MLOps team required.
FAQ
What is AI meeting transcription?
It turns meeting recordings into a transcript automatically, identifies who said what, and writes a summary and action items. MosaMinutes lets you edit while you listen in the same screen, until it's ready to send.
Can it transcribe recordings we already have?
Yes. Besides live recording, you can upload existing audio files; uploads go through a batch queue that isn't limited by the number of live streams. The transcript is punctuated, split into paragraphs and labeled by speaker, with a summary and action items.
Can meeting audio stay completely off the cloud?
Yes — that's the default, not an add-on. Recording, transcription, summaries and action items all happen on your own servers. It's containerized, runs on a single machine, and works on air-gapped networks.
Does it handle meetings in Taigi and Hakka?
Mandarin, Taigi, Hakka, Indigenous languages and English are all transcribed into Traditional Chinese with automatic punctuation. Hakka covers all six accents: Sixian, Hailu, Dapu, Raoping, Zhao'an and South Sixian.
Can it connect directly to Google Meet, Microsoft Teams or Zoom?
Not yet. Today you record or upload audio. Capturing online-meeting audio directly from your computer (without a bot joining the call) is on the roadmap, not a deliverable feature.
How many meetings can one machine transcribe at once?
One NVIDIA DGX Spark GB10 handles 10 live transcription streams; uploads go through a batch queue with no stream limit. Beyond one machine's capacity, it scales out to multiple machines.
If several people proofread at once, will edits get overwritten?
No. Proofreading runs on four independent channels that never overwrite each other. When there's a real conflict, the system doesn't silently pick a side — it flags the exact sentence.
Test it with your own meeting recordings.
Request a PoC
Product 03 · On-Premise
MosaScoreOn-prem call analytics Validating
Customer-service and telesales calls usually get 1–3% spot-checked. MosaScore hears every one in full and tags each utterance with emotion and intent.
Current status MosaScore can be deployed on standard Linux GPU servers (with an Apple Silicon single-machine demo), but it is not yet in production at a customer site, and the purchase-intent scores haven't been validated against real sales outcomes. We publish no accuracy figures.
PURCHASE INTENTAsked about payment — top-band anchor hit
84- CallerOh… I'm not really taking this stuff.22
- CallerI don't sleep well — I toss and turn.41
- CallerSo how many bottles come in a set?63
- CallerIf I buy, can I pay on delivery?84
01Coverage
From 1–3% spot checks
to every single call.
This is a difference in the order of magnitude of coverage — not a percentage gain in efficiency.
- 01Dual-channel separationSpeaker labels are 100% correct by design.
- 02One dictionary per industrySwap the dictionary, not the code.
- 03Every score has a reasonEmotion plus a 0–100 purchase-intent score, per utterance.
- 04Every call is checkableUtterance-level playback plus per-channel subtitles.
In developmentCompliance scorecardMono / mixed audioBulk QATaigi recognitionReal-time alerts
02On score validity
So far we've proven the pipeline works — not that the scores predict sales.The scores have not yet been validated against real sales outcomes, so we publish no accuracy figures. Validation is the first phase of any deployment: we test against your own recordings and actual results — and if the answer is negative, you'll know before you invest.
Field-testedEvery default backed by a test
Including the assumptions we had to throw out.
Most design decisions came from corrections after testing, not from a spec sheet. A QA system is only credible if its own development holds up to QA.
| Decision | Evidence |
|---|---|
| Switched the default recognition engine | Short-utterance latency dropped from about 2.5 s to about 0.5 s, with better recognition of domain terms |
| Dropped an even faster engine | It was twice as fast but produced wrong transcripts that "read like real speech" — more dangerous than obvious gibberish |
| Turned off the model's reasoning mode | The reasoning text was never read by the program, yet it pushed a single scoring pass from seconds to over 100 seconds |
| Don't let the AI rewrite transcripts | In testing it fixed one typo but over-edited two other sentences and deleted their endings |
| Agent turns aren't scored by default | Across six test calls, 61% of scoring compute went to content the interface never showed |
FitWhere it fits — and the cloud alternative
For organizations with lots of call recordings that can't use cloud analytics.
- Health-product and consumer telesalesPerformance analysis, script comparison and compliance checks on advertising claims. Already fully tuned for this scenario.Tuned
- Banking and insurance telesalesAlready required to record every call and keep it long-term; the compliance-audit need is the clearest.Compliance audit
- Contact-center QAFrom sampling to full coverage, and a source of material for complaint prevention and training.Full coverage
- Healthcare and long-term-care hotlinesData stays on the hospital's servers, with clear checks on what the agent may and may not say.Data stays on-site
Needs separate evaluation B2B sales: deals close over months, so a single call's "purchase probability" doesn't apply and the scoring dimensions would need to be redefined.
How it differs from cloud call analytics
If recordings leaving your premises isn't a concern and no regulation prevents it, cloud subscriptions are easier to adopt and currently more feature-complete.
| Cloud call analytics | MOSASCORE | |
|---|---|---|
| Recordings and transcripts | Uploaded to the vendor's servers | Stay entirely in your server room |
| Air-gapped networks | No — needs internet access | Yes — no internet needed |
| Pricing | Subscription by hours or call volume | Built once, no per-call API fees |
| Domain vocabulary | Mostly one shared, generic model | A dedicated dictionary per customer |
| Line-by-line verification | Depends on the vendor | Utterance playback + per-channel subtitles + full logs |
| Setup cost | Low — sign up and go | Needs servers and setup; we can supply hardware and install on site |
| Maturity | Commercial service, feature-complete | Core analysis pipeline complete; enterprise features in development |
PrinciplesA line we drew ourselves
Emotion inference is for customers only
Scoring contact-center agents is workplace monitoring, and emotion inference is its most contested form — the EU has banned emotion recognition in the workplace. Our principle: emotion inference applies only to customers; agent scoring is based on auditable, objective behavior and is never the sole basis for performance reviews. We drew this line at the design stage, not after a customer asked.
Open standards, no lock-in
Built on open-source models and standard compatible interfaces, so models can be swapped or upgraded. Where government or financial customers need a model set that meets supply-chain origin rules, we evaluate and swap in open-weight models from other sources, then re-validate quality.
SpecsSpecs & limitations
- Audio format
- Dual-channel .mp3 / .wav only for now (one track for the customer, one for the agent). Mono and mixed recordings are on the roadmap.
- Status
- Deployable on standard Linux GPU servers, but not yet in production at a customer site; purchase-intent scores are not yet validated against real sales outcomes.
- Processing speed
- Every utterance runs full recognition, emotion and scoring, so long calls take longer to process than their own length. Batch QA is unaffected; real-time in-call assistance needs a different architecture.
- Recognition tuning
- Brand names, product names and terms are passed to the recognizer as hot words. In testing this fixed product-name errors that no amount of after-the-fact text correction could recover — they can only be solved at recognition time.
- Post-processing
- Ten layers of text cleanup, including simplified-to-traditional conversion, hallucination and repetition filtering, filler removal, punctuation and number normalization, and homophone correction with pinyin-similarity matching.
- Scoring
- A local LLM combines the transcript, conversation context and acoustic emotion to output emotion, intent, a 0–100 score and the reason for any change after each customer utterance; calibrated with five anchored bands and example sentences, and unaffected by the agent's script.
- Server deployment
- Ubuntu Linux x86_64 + NVIDIA GPU, fully functional on an intranet; browser-based, nothing to install.
- Laptop demo
- Apple Silicon M-series with 16 GB+ unified memory (24 GB recommended). Model weights need internet once to download; after that analysis runs fully offline, and the front end has no external dependencies — it demos in a room with no network.
FAQ
What can call quality analytics actually do?
Customer-service and telesales calls that used to be 1–3% spot-checked become searchable data: every call heard in full, transcribed line by line, tagged with emotion and a 0–100 purchase-intent score, with a reason for every score change. It currently requires dual-channel recordings.
How accurate are the scores?
The scores haven't been validated against real sales outcomes, so we publish no accuracy figures. Validation is the first phase of deployment: we test against your own recordings and actual results before anything else.
Do recordings need to go to the cloud?
No. The system runs on a Linux GPU server in your server room; recordings and transcripts never leave it.
How is this different from a manager spot-checking calls?
Spot checks see a sample; this sees every call. And every judgment is checkable on the spot: utterance-level playback plus per-channel subtitles, with dual-channel separation so speaker labels are correct without guessing.
Every industry talks differently. Does that require code changes?
No. Each industry gets its own domain dictionary — swap the dictionary, not the code.
On-Premises
On-prem speech AIFour models, all running on your own servers
No calls to any cloud AI service. Containerized, runs on a single machine, and works on air-gapped networks.
- MODEL 1Speech recognition
- MODEL 2Diarization
- MODEL 3Dialogue
- MODEL 4Speech synthesis
- Audio
- Transcripts
- Summaries
- Voiceprints
- Domain dictionaries
Not a single clip of audio, transcript or prompt goes out.
MosaTalk's edge-split mode also needs one small cloud host; standalone mode needs only your intranet.
01When it has to be on-prem
- 1The content itself is confidentialBoard decisions, compliance reviews, client interviews, medical records.
- 2The network is already closedOn an air-gapped network, anything that needs the internet is off the table.
- 3Procurement specifies the sourceOn-prem lets you specify exactly where the models come from.
On-prem isn't a selling point. For some organizations, it's a prerequisite.If none of the three cases above applies to you, we'll tell you straight: the cloud is a better fit.
02Cloud vs. on-prem
We list the row
where the cloud wins, too.
| Cloud speech services | VOXMOSA | |
|---|---|---|
| Audio and transcripts | Uploaded to the vendor's servers | Stay entirely in your server room |
| Air-gapped networks | No — needs internet access | Yes — no internet needed |
| Taigi / Hakka / Indigenous | Local services exist, all cloud-based | Supported, running on-prem |
| Pricing | Subscription — the more you use, the more you pay | Built once, no per-call API fees |
| Model customization | No — one shared, generic model | Yes — we train our own models |
| Voiceprints & dictionaries | Usually can't be taken with you | Stored on your side, with the system |
| Setup cost | Low — sign up and go | Needs servers and setup; we can supply hardware and install on site |
03Can one machine handle it?
Live meeting transcription on one NVIDIA DGX Spark GB10
The full voice-agent pipeline, validated on a single 46 GB card
Recognition, diarization, dialogue and synthesis — all inside your server room
These numbers come from environments we've run ourselves, not from a spec sheet; beyond one machine's capacity, it scales out to multiple machines. We can supply the hardware and install it on site.
SizingConfigurations we've validated
| Item | MosaMinutes (transcription) | MosaTalk (voice agent) |
|---|---|---|
| Machine | One NVIDIA DGX Spark GB10 (128 GB unified memory) | One NVIDIA GPU with plenty of memory (validated on a 46 GB card) |
| Operating system | Stock DGX OS (Ubuntu-based) | Ubuntu Linux x86_64 |
| Deployment | Containerized, runs on a single machine | Standalone mode needs only your intranet |
| Clients | Any modern browser, nothing to install | WebRTC browsers, or dial in from a phone |
| Hardware supply | We can supply and install on site | We can supply and install on site |
PlannerDeployment planner
Run the numbers yourself first
Pick a product, languages and scale to see the verified configuration, and what this combination can and cannot do today. Every number comes from the tested setups above.
MosaMinutes
- Host
- NVIDIA DGX Spark GB10 (128 GB unified memory), stock DGX OS
- How many
- 1 unit: one unit transcribes 10 meetings live, measured
- Languages
- Mandarin: SupportedTaigi: Supported
- Network
- Containerized, runs on a single machine, no outbound access needed
- Know first
- Cannot join Google Meet, Teams or Zoom directly yet; record or upload audio instead
IndustriesIndustries & compliance
For these organizations, the use case is the compliance requirement.
We tailor access control, audit logs and data retention to each industry's actual regulations and internal controls, rather than applying one template.
- GovernmentOfficial meetings, public hearings and committee records. Running on-prem meets both security rules and data-residency requirements.Official document exportplanned · supply-chain rules
- Financial servicesBoard, compliance and risk meetings, plus high-volume product inquiries.Internal control · audit trail
- HealthcareInternal meetings, case reviews, patient education and appointment inquiries, with data kept on hospital servers.Medical records · sensitive-field maskingplanned
- Law firmsClient interviews and case discussions. Confidentiality duties make on-prem a requirement, not an option.Confidentiality · evidentiary records
- Long-term care, education & researchFamily consultations, service briefings, interviews and teaching records, plus preservation of local-language speech data.Indigenous-language preservation · interview transcripts
- EnterprisesRegular meetings, project reviews, client meetings and support-line routing — anywhere content must not leak.SSO · retention settings
- OIDC single sign-on
- Tiered permissions
- Audit logs
- Result retention
- Containerized
- Air-gap friendly
- Model-origin options
- Duplicate-request protection
After deployment
The hard part of on-prem isn't installing it — it's who maintains it, how it meets internal controls, and how long data is kept. All of that is part of the engagement.
More than delivering software
Scenario setup, industry compliance adjustments, hardware supply and on-site installation, maintenance and version updates, and evaluation reports that meet academic standards.
Compliance is tailored, not templated
Government, financial, healthcare and legal requirements differ. We adjust access control, audit logs and data retention to your actual regulations and internal controls.
FAQFAQ
What is on-prem AI? Is it the same as private cloud or local deployment?
In Taiwan's procurement language, on-prem, local, private and intranet deployment usually mean the same thing: models and data live on machines your organization controls, not in a vendor's cloud. "Private cloud" is broader — sometimes your own data center, sometimes a rented dedicated cloud environment. Ours is the former: physical machines on your premises, with no internet required.
Does it work on networks with no internet access at all?
Yes. It's containerized and runs on a single machine, completing recognition, summaries and speech synthesis without going online. The exception is MosaTalk's edge-split mode, which needs one small cloud host; standalone mode needs only your intranet.
Is on-prem always more expensive?
The upfront cost is higher — cloud services work as soon as you sign up, while on-prem needs servers and setup. The difference comes later: cloud subscriptions charge by seat or by hour, so the more you use the more you pay; on-prem is built once with no per-call API fees. The answer depends on your volume and how many years you'll use it, not the first invoice. We can run the numbers with your actual meeting or call volume.
Can the models be tuned to our terminology?
Yes. Our speech recognition models are trained on architectures we built ourselves rather than by adjusting a generic model's parameters, so they can be deeply customized for your domain terms, recording conditions and acoustics. MosaScore also uses one domain dictionary per industry — swap the dictionary, no code changes.
Our procurement rules specify model origin. Can you meet that?
Yes. Model sources and the supply chain can be specified to match your procurement rules. That's a structural difference from cloud services, where the buyer usually has no control over the vendor's model stack.
Is keeping audio off the cloud a legal requirement, or an internal policy?
We've seen both, and they come from different places. Personal data and medical records are governed by Taiwan's Personal Data Protection Act; government agencies also have the Cyber Security Management Act and its security-level requirements; financial institutions follow their regulator's internal-control and audit rules. Even more common is a third case: no law forbids the cloud, but internal policy or a client contract says data may not leave your environment. We don't decide which applies to you — that's for your compliance and security teams. Our job is to make "not in the cloud" technically real, and to adjust access control, audit logs and retention to your actual requirements.
What is your recognition accuracy?
We don't give marketing numbers. Character error rates, competitor comparisons, test methodology and raw audio will be published in full to academic standards, and we can evaluate on public benchmarks you choose or your own audio. Until then, the most meaningful number is the one measured on your own recordings in a proof of concept.
Taiwan Languages
Taiwan's languagesTaigi, all six Hakka accents, and Indigenous languages
Our speech recognition models are trained on architectures we built ourselves — not by tweaking an off-the-shelf model.
- Mandarin
- TaigiCode-switching OK
- HakkaAll six accents
- Indigenous16 peoples · 42 languages
- English
Which product supports what
| Language | Meeting minutes | Voice agent |
|---|---|---|
| Mandarin | Supported | Listen & speak |
| Taigi | Supported | Understands; speech planned |
| Hakka | Supported | Planned |
| Indigenous | Supported | Not yet |
| English | Supported | Listen & speak |
Six Hakka accents Recognition · Synthesis
- Sixian
- Hailu
- Dapu
- Raoping
- Zhao'an
- South Sixian
Tap an accent to listen. All six clips say the same sentence: "Hello everyone, we are a speech company." They are synthesized speech, not human recordings.
「客語語音合成、語音識別技術由牧仁資訊研發」
"Hakka speech synthesis and recognition technology developed by Voxmosa (牧仁資訊)."
From the Gohakka platform's copyright notice. Don't take our word for it — go look. Gohakka's machine translation and language models are provided by the platform operator, Shunxi Technology (瞬息科技), not by us; the Hakka corpus belongs to the Hakka Affairs Council, and our team handled part of its preparation.
Code-switching is how Taiwan actually talks.Few people in Taiwan's meetings or call centers speak pure Taigi for long. In practice it's a sentence of Mandarin, a sentence of Taigi, with a few English words in between.
Our Indigenous-language models are built in a long-term partnership with iThuan, deployed in the Indigenous Languages Research and Development Foundation's language AI system, and published on Hugging Face. The corpus belongs to the Council of Indigenous Peoples, and our team handled part of its preparation. Peoples and language counts follow the Council of Indigenous Peoples' classification, and Hakka accents follow the Hakka Affairs Council's six-accent classification — both can be checked item by item. Competition results and publications
FAQFAQ
Can Taigi recordings be transcribed?
Yes, including recordings that mix Taigi and Mandarin. MosaMinutes transcribes Taigi into a Traditional Chinese transcript with automatic punctuation and speaker labels, all on your own machines.
Which Hakka accents does transcription support?
All six: Sixian, Hailu, Dapu, Raoping, Zhao'an and South Sixian, for both recognition and synthesis. The classification follows the Hakka Affairs Council's six accents and can be checked item by item.
What about Indigenous languages?
MosaMinutes supports 42 languages across 16 Indigenous peoples. Our team built these models in a long-term partnership with iThuan, and they're public on Hugging Face for anyone to download and check. The voice agent doesn't support Indigenous languages yet.
How is this different from a general-purpose speech recognition service?
General services usually add local-language support by fine-tuning an existing model, and accuracy drops as accents and code-switching pile up. Our models are trained on architectures we built ourselves, so they can be customized for specific accents, terminology and recording conditions. The other difference is deployment: local-language recordings are often the most sensitive, and general services all run in the cloud.
What is the recognition accuracy?
We don't give marketing numbers. Accuracy for local languages depends heavily on accent, recording conditions and how much code-switching there is, so a percentage without a test set means nothing. Character error rates, competitor comparisons, methodology and raw audio will be published in full; until then, the most meaningful number is the one measured on your own recordings.
Research-ledTeam & research
A research-led speech team.
Long focused on speech recognition for Taiwan's languages. Our ASR models are trained on architectures we built ourselves, and our open-source project FormoSpeech is on Hugging Face.
- 2025Best Technical Paper AwardO-COCOSDA 2025
- 20221st place, Hakka ASR Challenge (general track)FSR 2022
- 20201st place, Taiwanese ASR Challenge (industry track)FSR 2020
- 20182nd place, Mandarin ASR Challenge (industry track)FSR 2018
- 2008Best Student Paper AwardISCSLP 2008
Publications 41 papers
- 2026LRECTG-ASR: Translation-guided learning with parallel gated cross attention for low-resource automatic speech recognitionarXiv ↗
- 2026LRECEfficient dialect-aware and conditioning for low-resource Taiwanese Hakka speech processingarXiv ↗
- 2026TASLPUniversal robust speech adaptation for cross-domain speech recognition and enhancementarXiv ↗
- 2026ICASSPSCENE: Semantic-aware codec enhancement with neural embeddingsarXiv ↗
- 2026InterspeechPersonalized keyword spotting for user-defined keywords leveraging text-independent speaker verificationarXiv ↗
- 2026IEEE ICMERobust generative audio quality assessment: Disentangling quality from spurious correlationsarXiv ↗
- 2025O-COCOSDAExploring the impact of data quantity on ASR in extremely low-resource languagesarXiv ↗
- 2025IEEE ASRURevealing the role of audio channels in ASR performance degradationarXiv ↗
- 2025ICASSPChannel-aware domain-adaptive generative adversarial network for robust speech recognitionarXiv ↗
- 2025IEEE ASRUQAMRO: Quality-aware adaptive margin ranking optimization for human-aligned assessment of audio generation systemsarXiv ↗
- 2025APSIPA ASCDRASP: A dual-resolution attentive statistics pooling framework for automatic MOS predictionarXiv ↗
- 2024O-COCOSDAVoxHakka: A dialectally diverse multi-speaker text-to-speech system for Taiwanese HakkaarXiv ↗
- 2024IEEE SLTEffective noise-aware data simulation for domain-adaptive speech enhancement leveraging dynamic stochastic perturbationarXiv ↗
- 2023InterspeechA training and inference strategy using noisy and enhanced speech as target for speech enhancement without clean speecharXiv ↗
- 2023IEEE SPLMulti-target filter and detector for unknown-number speaker diarizationarXiv ↗
- 2022InterspeechChain-based discriminative autoencoders for speech recognitionarXiv ↗
- 2022ISCSLPSpeech-enhanced and noise-aware acoustic modeling for robust speech recognitionarXiv ↗
- 2022EUSIPCOFilter-based discriminative autoencoders for children speech recognitionarXiv ↗
- 2022InterspeechDisentangling the impacts of language and channel variability on speech separation networksarXiv ↗
- 2021InterspeechAlloST: Low-resource speech translation without source transcriptionarXiv ↗
- 2021CSRCThe Academia Sinica system for the SLT 2021 children speech recognition challenge
- 2021InterspeechDual-path filter network: Speaker-aware modeling for speech separationarXiv ↗
- 2021APSIPA ASCGeneration of speaker representations using heterogeneous training batch assemblyarXiv ↗
- 2020O-COCOSDAUsing Taigi dramas with Mandarin Chinese subtitles to improve Taigi speech recognition
- 2020CHiMEThe Academia Sinica systems of speech recognition and speaker diarization for the CHiME-6 challenge
- 2020TASLPSubspace-based representation and learning for phonotactic spoken language recognitionarXiv ↗
- 2019InterspeechExploring the encoder layers of discriminative autoencoders for LVCSR
- 2019O-COCOSDAMandarin mispronunciation detection and diagnosis feedback using articulatory attributes based multi-task learning
- 2019APSIPA ASCMulti-task learning for acoustic modeling using articulatory attributes
- 2017InterspeechDiscriminative autoencoders for acoustic modeling
- 2017IJCLCLPA replay spoofing detection system based on discriminative autoencoders
- 2017ICASSPDiscriminative autoencoders for speaker verification
- 2017ROCLINGSpeaker diarization based on i-vector PLDA scoring and using GMM-HMM forced alignment
- 2016InterspeechMinimization of regression and ranking losses with shallow neural networks on automatic sincerity evaluation
- 2014InterspeechClustering-based i-vector formulation for speaker recognition
- 2014ICASSPSpeaker verification using kernel-based binary classifiers with binary operation derived features
- 2014InterspeechEnsemble of machine learning algorithms for cognitive and physical speaker load detection
- 2013ICASSPSubspace-based phonotactic language recognition using multivariate dynamic linear models
- 2012InterspeechSubspace-based feature representation and learning for language recognition
- 2009IEEE ASRUGeneralized likelihood ratio discriminant analysis
- 2007IEEE ASRUTraining data selection for improving discriminative training of acoustic models
Press In Chinese
- Hakka fairy tales and two-host podcasts in one second: inside "Hakka GPT"Hakka News
- Tackling two major challenges in AI speech technologyEconomic Daily News
- Real-time speech enhancement aimed at the streaming industryCommercial Times
- VoxHakka, the first Hakka text-to-speech system, launchesEconomic Daily News
- Cross-domain speech enhancement shows AI innovationEconomic Daily News
Headlines translated from Chinese. Coverage of technology led by core team members at previous organizations.
NotesResearch notes · in preparation
Technical articles, evaluation methodology and observations on Taiwanese speech data — where we'll spell out what "academic rigor" means in practice.
- Coming soonHow to evaluate Taigi speech recognition fairly
- Coming soonFitting three models into one GPU's memory
- Coming soonWhy meeting proofreading shouldn't be "last write wins"
PoCProof of concept
Prove it on your own audio first.
Then we talk deployment.
We run a proof of concept on your real meeting recordings or customer calls, and deliver a reproducible evaluation report with a clear methodology.
Can it run entirely in our own server room?
Yes — that's the default, not an add-on. It's containerized, runs on a single machine, and works on air-gapped networks with no internet access.
What hardware do we need?
One NVIDIA GPU with plenty of memory runs the whole system. MosaMinutes handles 10 live streams on one NVIDIA DGX Spark GB10; MosaTalk is validated on a single 46 GB card. We can supply the hardware and install it on site.
What's your recognition accuracy?
We don't give marketing numbers. Character error rates, competitor comparisons, test methodology and raw audio will all be published in full. Until then, the number that matters is the one measured on your own recordings.
How is this different from a cloud service?
Recordings and transcripts stay entirely in your server room, it works air-gapped, and there are no per-call API fees. But we don't claim it's better across the board — without compliance or security constraints, a cloud subscription is usually cheaper and more flexible.