Mount Software

Our models

Four models we built

Each model below is fine-tuned for Romanian and measured on a named test set.

STT

Whispro speech-to-text

Romanian transcription that copes with English tech terms and phone audio.

8.62% WER

vs 17.69% for stock Whisper, on noisy Romanian-English tech speech

See Whispro
TTS

Custom text-to-speech

Natural Romanian voices, fine-tuned for your use case.

About $0.0006

per 1,000 characters of compute

See custom TTS
EMB

Romanian embeddings

mount-embeddings-0.6b, for search and retrieval over Romanian text.

0.683 nDCG@10

vs 0.577 for bge-m3, on Romanian legal retrieval

See embeddings
IE

Data extraction

A 4B-parameter model that turns business documents into structured fields.

0.958 F1

on noisy scanner data, up from 0.851 before fine-tuning

See data extraction

Whispro, speech-to-text

Romanian speech recognition built for tech terms and phone calls

Whispro is a Whisper model we fine-tuned for Romanian, with a focus on English words inside Romanian sentences and on 8 kHz phone audio.

Fewest errors of the four systems tested Noisy Romanian-English tech speech WER %, **lower is better**
System	WER %
Whispro	8.62
ElevenLabs Scribe v2	11.33
OpenAI gpt-transcribe	13.45
Stock Whisper large-v3	17.69

Fewest errors on real call-centre calls Real call-centre calls, 59 calls WER %, **lower is better**
System	WER %
Whispro	18.86
Gemini 3.1 Pro	19.43
OpenAI gpt-transcribe	22.08

Real phone calls vs stock Whisper

18.73 to 15.64 WER %

WER falls on 43 real phone calls.

Romanian with English terms.
Product names, code and API terms in the middle of a Romanian sentence.
Phone audio.
Tuned on 8 kHz call recordings, not only on clean studio speech.
Your hardware.
Runs on one GPU or on CPU servers, so audio stays where you put it.
Your domain, on request.
We fine-tune on your vocabulary, for example medical terms or your own product names, and measure on your audio before we promise anything.

Cost per audio minute

Whispro costs about $0.001 per audio minute (about $0.06 per hour) in GPU compute on a Cloud Run L4.

About 3-6x cheaper than the major speech APIs, and about 15x cheaper than Google STT standard.

Whispro compute cost vs API list prices Per audio minute, **lower is better**
Service	List price per audio minute	Multiple of Whispro cost
Whispro (compute only)	about $0.001	1x
AssemblyAI Universal-3.5 Pro	$0.0035	3.3x
ElevenLabs Scribe v2	$0.00367	3.5x
Deepgram Nova-3 (monolingual)	$0.0043	4.1x
OpenAI gpt-transcribe	$0.0045	4.3x
OpenAI gpt-4o-transcribe	$0.006	5.7x
Google STT V2 Standard	$0.016	15.3x

Test Whispro on your audio

Custom text-to-speech

Romanian voices tuned to your use case, at a fraction of API cost

We fine-tune a compact voice model on text from your use case, such as clinic reception, food ordering or a phone agent, and run it on a single GPU.

50-140x below ElevenLabs, 20-50x below OpenAI, Google and Azure Compute cost vs list price, per 1,000 characters US dollars, **lower is better**
Service	Price per 1,000 characters
Our custom voice (compute only)	about $0.0006
OpenAI tts-1	$0.015
Azure Neural	$0.015
OpenAI tts-1-hd	$0.030
Google Chirp 3 HD	$0.030
ElevenLabs Flash / Turbo	$0.04
ElevenLabs Multilingual v2 / v3	$0.08

Quality

UTMOS of 4.13 for our voice vs 3.75 for the hosted reference voice, scored by an automatic predictor.

Speed

First audio in 78-140 ms, measured server-side.

Tuned per use case.
The voice is fine-tuned on text from your domain, so the vocabulary and phrasing match your calls.
One GPU.
Runs on a single Cloud Run L4.
A CPU fallback.
A lighter voice runs on CPU when no GPU is available; its quality is lower.

Hear a voice for your use case

Romanian embeddings

mount-embeddings-0.6b: search that understands Romanian

mount-embeddings-0.6b is fine-tuned from bge-m3 for Romanian retrieval.

Ranks the right passage higher than bge-m3 and OpenAI nDCG@10 on Romanian legal retrieval Scale 0 to 1, **higher is better**
Model	nDCG@10
mount-embeddings-0.6b	0.683
bge-m3	0.577
OpenAI text-embedding-3-large	0.553
BM25 (keyword search)	0.440

Fine-tuned for your use case

We fine-tune the model for your domain and measure it on a sample of your documents before rollout.

Retrieval for RAG and search.
Find the right passage in contracts, regulations and internal documents.
Your documents, your index.
The model runs on your infrastructure, so your text is not sent to an embedding API.
Retrained on your corpus.
We fine-tune on your documents when the general model is not enough.

Evaluate embeddings on your documents

Data extraction

Structured data from business documents, with a 4B-parameter model

We fine-tune a 4B-parameter open model to pull the fields you need out of business documents and return valid JSON.

3.5x fewer errors on noisy scanner data Share of fields missed or wrong (1 − F1), lower is better Untuned 4B model vs the same model fine-tuned · bars start at 0%
Model	Fields missed or wrong (1 − F1)	F1
Clean
Untuned 4B model	10.2%	F1 0.898
Fine-tuned	3.5%	F1 0.965
Noisy scanner data
Untuned 4B model	14.9%	F1 0.851
Fine-tuned	4.2%	F1 0.958

Against API models

Against the best API model we tested, the fine-tuned model is ahead by 1.9 F1 points on clean documents and 4.6 points on noisy ones.

Reliability

Valid JSON on all 1,473 test documents.

Your fields, your schema.
We define the field list with you and train on it.
Steady on noisy scans.
F1 drops by less than one point from clean to noisy text.
One GPU, your infrastructure.
Runs on a single L4-class GPU; on-premises deployment is possible.
Text in, JSON out.
It reads the OCR or text layer, so it pairs with the OCR you already use.

Send us a document sample

Custom models

Custom models trained on your own data

A small model trained on your data and your metric can beat a general model on one narrow task, and it runs on hardware you control. This is how we get there, and the four models above are our proof points.

Define the task and the metric.
We agree the task, the output and the success measure with your experts.
Build an eval set from your data.
A fixed, labelled test set from your own records, kept apart from training.
Baseline frontier models.
We score the leading API models on that set, so you see what you have today.
Curate training data.
Your data, plus generated examples where real data is thin.
Fine-tune.
Supervised fine-tuning first, then reinforcement learning with verifiable rewards where it pays.
Gate on evals.
A model ships only if it beats the baseline on your eval set.
Deploy a small model.
On your infrastructure or in an EU cloud, behind your existing application.
Monitor and retrain.
We track drift, refresh the eval set and retrain when it moves.

Proof points from our own models

Whispro
WER 8.62 vs 17.69 for stock Whisper on noisy tech speech; 15.64 vs 18.73 on real phone calls.
Embeddings
nDCG@10 of 0.683 vs 0.577 for bge-m3.
Data extraction
F1 from 0.851 to 0.958 on noisy scanner data, from a 4B-parameter model.
Text-to-speech
UTMOS of 4.13 vs 3.75 for the hosted reference voice.

Not every task needs a custom model. If a general API already clears your eval set, we say so.

Book a model consultation How we run AI projects

AI phone agents

Romanian phone agents on our own speech models

Our phone agents listen with Whispro and speak with our custom Romanian voice, and they book, answer and hand off in a single call. Compute cost is about 4 cents per call-minute, excluding telephony.

Dental and clinic appointments.

Book, cancel and reschedule; look up the caller's appointments; answer clinic questions on prices and required documents; send a confirmation SMS.
Restaurant food ordering.

Take the order on the same cart as the website, recognise the caller by phone number, and pass the call to a person when needed.
Reception and front desk.

Answer the line, take callbacks and notes, and transfer to a person during opening hours.
Support hotline.

Answer common questions from your FAQ, collect the details, and hand the case to your team.

Romanian-native.
Built on our own speech-to-text and text-to-speech.
Checked after the call.
An automatic QA pass reviews calls after they end.
Your systems.
A shared booking core connects to clinic scheduling systems.

Talk about a phone agent

Software engineering

Custom software development

We deliver quality and cost effective technical solutions that provide you with de-risked, quality products.

Start your project

About Us

Software product development is in our dna

How we differentiate is in our experience and mindset. Our company did not evolve out of IT Consulting or a digital agency.

What makes us unique is the combination of technical excellence with an entrepreneurial mindset.We are a team comprised of tech entrepreneurs and product development people, with our focus on engineering excellence and focused on value creation through software.

We understand what it requires to build impactful and successful digital experiences cutting through the noise, and providing solutions that are purpose-driven and results-oriented.

See About