Speech and Transcription Business Due Diligence

Due diligence on a speech-to-text or voice-analytics business is largely a paper trail exercise, because the risk in this kind of company sits in what data trained the models and who actually owns them, not in whether the technology works. A buyer’s technical team can test accuracy in an afternoon. Establishing whether every recording behind that accuracy was lawfully collected, and whether the models built from it are actually owned free and clear, takes considerably longer, and it is where most of the real findings in this sub-sector come from.

Documents to request

• Consent records for every voice dataset used in training, organized by dataset rather than summarized as a single policy statement.
• Data-processing agreements with each enterprise customer, and confirmation they match what the system actually does in practice.
• Contractor and employment agreements covering anyone who helped build an acoustic or language model, with the intellectual-property assignment clause specifically confirmed rather than assumed.
• The actual agreement with any third-party speech-model vendor the product depends on, including pricing terms and what happens if that vendor changes price or shuts off access.
• Independent accuracy benchmark results if any exist, separate from anything the company has published in its own marketing.

Registry and status searches worth running

A corporate status and good-standing check confirms the selling entity is actually in a position to complete the transaction it is proposing. An execution and judgment search surfaces any outstanding claims against the company that could attach to assets being purchased. And a search at the Canadian Intellectual Property Office for any patents or trademarks actually filed on the technology tells a buyer whether the proprietary claims made in the sales process have anything registered behind them — many businesses in this space have filed nothing at all, which is itself a useful and fairly common finding, not automatically a red flag on its own.

Confirm the technical integrations actually transfer

A transcription or voice-analytics product is rarely self-contained — it typically depends on integration agreements with telephony platforms, contact-centre software or practice-management systems to capture audio automatically, and those agreements deserve the same scrutiny as the customer contracts themselves. Request the actual integration or partner agreement for each platform the product connects to, not a description of the connection, and check specifically whether it includes a change-of-control clause requiring the partner’s consent before the integration continues to function under new ownership. Confirm which integrations run against a current, supported version of the partner’s API and which are legacy connections that could break without warning, since a business that looks technically solid on paper can still depend on a handful of ageing, undocumented connections that only the departing owner fully understands. Where an integration partner is itself a potential competitor or has its own transcription ambitions, also check whether the partnership agreement includes any exclusivity or non-compete language that could constrain what the buyer is allowed to build after closing.

Findings that kill deals

• Training audio with no documented speaker consent, including scraped or purchased datasets whose origin cannot actually be traced.
• Voiceprints retained past the purpose for which they were originally collected, with no documented basis for the extended retention.
• No intellectual-property assignment on record for a model built by a contractor who has since left the company.
• An undisclosed dependency on a single speech-model vendor whose pricing or availability could change with little warning to the business.

What a finding actually means

An undocumented consent gap is not a paperwork nuisance to be cleaned up after closing — it is a liability that transfers with the company, because the buyer becomes the party legally responsible for personal information it cannot demonstrate was lawfully collected in the first place. An unassigned model built by a departed contractor is a more direct problem still: it can mean the buyer is not actually acquiring clear ownership of the core technology asset it is paying for, and the original builder retains a claim on it. Neither finding necessarily kills a deal by itself, but both change the price, the structure, or the representations and warranties a seller needs to give before a buyer’s counsel will let the transaction proceed.

Reading the customer contracts closely

Confirm whether each enterprise contract can actually be assigned to a new owner automatically, or whether assignment requires the customer’s active consent — this single point determines whether the revenue base a buyer is paying for actually survives the transaction. A contract silent on assignment is not necessarily assignable by default, and a customer who declines to consent after closing can remove a meaningful share of the acquired revenue with no recourse for the buyer beyond whatever the purchase agreement anticipated.