Voice Data for AI | Voices & Media Solutions

Voice data collection for AI model training: quality, diversity and annotation rigour

Voice data is the raw material of artificial intelligence systems that process, recognise and synthesise human speech. The quality of an automatic speech recognition model, a text to speech system or a voice-enabled virtual assistant depends directly on the quality, diversity and annotation rigour of the data it was trained on. Without good data, there is no good model.
At Voices & Media Solutions, we supply voice data to technology companies, telecommunications businesses, artificial intelligence startups and research teams developing or improving vocal processing systems. Our unique position in the Portuguese language market — with coverage of Portugal, Brazil, Angola, Cape Verde and Mozambique — makes us a difficult partner to replace for those who need data representative of Portuguese variants. We also work with native voices in over 70 languages.

Bespoke voice data collection with speaker profiles defined by the client.

Coverage of all Portuguese variants: European, Brazilian, Angolan, Cape Verdean and Mozambican.

Native voices in over 70 languages for multilingual projects.

Transcription included in the process.

Audio quality control under controlled conditions with professional equipment.

ISO 9001 certified company.

The raw material of AI voice systems

Voice data consists of sets of human speech recordings that are organised, transcribed and annotated for use in training artificial intelligence models. These are not random recordings or low-quality captures. A voice dataset usable for AI training must meet a rigorous set of technical and linguistic criteria: controlled audio quality, voice diversity, coverage of speech contexts and styles, accurate transcriptions and annotations that enable the model to learn relevant patterns.
This data feeds three main categories of systems:

Automatic Speech Recognition (ASR): converts speech to text. Powers virtual assistants, automatic transcription systems and voice interfaces.
Text to Speech synthesis (TTS): converts text to speech naturally. Powers text to speech systems, screen readers and voice assistants.
Spoken Language Understanding (SLU): interprets the meaning of what is said. Powers dialogue systems and conversational agents.

Each of these applications has specific requirements in terms of the type of data needed, which makes bespoke collection frequently more efficient than relying on generic datasets.

Bespoke collection, aligned with model requirements

Our service focuses on bespoke voice data collection. When a client needs data with specific characteristics, we manage the entire process: defining the required speaker profiles, creating reading scripts or spontaneous speech scenarios, conducting recording sessions under controlled conditions, and transcribing and annotating the material produced.
Bespoke collection is the right solution when a project requires a precise demographic profile, a specific accent, a particular subject domain or a volume of data not available from other sources. It is a more time-intensive process than accessing pre-existing datasets, but it guarantees data fully aligned with the requirements of the model to be trained.

Unique coverage of Portuguese variants

Portuguese is the fifth most spoken language in the world, with over 260 million native speakers across three continents. Yet in the context of voice data for AI, it remains an under-represented language — particularly in its African variants. Most available datasets cover European Portuguese or Brazilian Portuguese reasonably well. Representative data for Angolan, Cape Verdean or Mozambican Portuguese is scarce.
Voices & Media Solutions has the capacity to supply voice data across all major Portuguese variants: European, Brazilian, Angolan, Cape Verdean and Mozambican. This coverage results from years of work with native speakers from each region and a network of voice professionals across Portuguese-speaking countries. For companies developing speech recognition or voice synthesis systems in Portuguese, this coverage is a critical resource.

What determines the real utility of a voice dataset

The quality of a voice dataset is not measured solely by the number of hours recorded. The criteria that determine the real utility of data for AI model training include:

Audio quality: recordings in acoustically controlled environments, with professional equipment and without background noise that could compromise signal clarity.
Speaker diversity: coverage of different genders, age ranges, regional accents and speech profiles to ensure model robustness.
Context coverage: data representing different speech styles — from text reading to spontaneous and conversational speech — according to model requirements.
Rigorous transcriptions: text aligned with audio at word or phoneme level, depending on the level of detail required.
Relevant annotations: metadata on the speaker, recording context, speech style and other variables that enhance the dataset's training value.

What type of projects use it

ASR model training: automatic speech recognition for virtual assistants, automatic transcription and voice interfaces.
TTS system development: voice synthesis with improved naturalness for any language and variant.
Academic and scientific research: computational linguistics, natural language processing and phonetics.
Evaluation and benchmarking: testing and comparing the performance of existing voice models.
Improving models in production: additional data targeting specific performance gaps identified in real usage.

Technical brief: definition of dataset requirements in terms of language, speaker profile, volume, delivery format and annotation level.

Proposal and validation: presentation of the recommended solution, with volume, timeline and cost estimates. For bespoke collection, this includes speaker profiles and proposed scripts.

Production or curation: bespoke data collection with recording sessions under controlled conditions, or selection and preparation of existing datasets, with transcription and annotation to specification.

Quality control: review of produced material prior to delivery, including audio quality verification, transcription accuracy and annotation completeness.

Delivery: supply of data in agreed formats, with technical dataset documentation and support for integration into the client's training pipeline.

Do you need voice data for your AI project?

Every project has different requirements. Some need volume. Others need linguistic specificity. Others need both. Our technical team is available to assess what you need and recommend the most efficient approach — whether bespoke collection, access to existing datasets or a combination of both.

Request contact

What is the difference between voice data for ASR and for TTS?

For automatic speech recognition (ASR), data needs to represent the real diversity of human speech: different accents, rhythms, contexts and acoustic conditions. The model must learn to recognise speech in varied conditions. For voice synthesis (TTS), data typically consists of recordings from one or a few voice artists under highly controlled conditions, with high audio quality and complete phonetic coverage. The goal is for the model to learn to replicate a specific voice naturally. In many respects, these are opposite requirements.

What is the minimum volume of data needed to train a model?

It depends on the type of model and the objective. To adapt a pre-existing model to a specific accent or domain, a few dozen hours may be sufficient. To train a model from scratch at production quality, hundreds or thousands of hours are typically needed. Our technical team helps define the appropriate volume based on the specific requirements of each project.

Can you supply voice data in languages other than Portuguese?

Yes. We work with native voice artists in over 70 languages. For multilingual projects or specific languages, contact us with your requirements and we will present a proposal with the available options.

Is transcription included in the data supply?

Yes. Transcription is part of our bespoke collection process. The level of detail — word-level, phoneme-level or with additional annotations — is defined in the initial brief and reflected in the proposal.

What formats are the data delivered in?

Audio files are typically delivered in WAV (uncompressed) or FLAC, with transcriptions in standard formats or other formats used by the main AI model training frameworks. The exact format is defined according to the client's training pipeline requirements.

Can you collect spontaneous speech data or only text reading?

Both. For spontaneous speech recognition systems, collection includes conversation scenarios, question and answer sessions and real usage contexts. For voice synthesis systems, collection is typically based on structured text reading with controlled phonetic coverage. The type of collection is defined according to model requirements.

How long does a voice data collection project take?

It depends on the volume, the number of speakers required and the annotation level. For small-scale projects, a few weeks. For large-volume projects with many languages or accents, the timeline is defined in the brief and may extend over several months. Our team always presents a detailed schedule in the proposal.

Clients