Gladia review: smart transcriptions, but not in all languages

The French transcription tool Gladia aims to distinguish itself from other tools with smart AI features. It works, depending on what language you speak.

Anyone who has ever had to manually transcribe an audio recording knows how time-consuming and unrewarding it can be. For journalistic work, a transcription tool is also no unnecessary luxury. It can save you many hours if you only have to upload a recording and get a written transcript with the click of a button.

ITdaily is testing several European transcription tools. After Good Tape from Denmark, it is now the turn of the French Gladia. With more than 300,000 users, this is no small player in the transcription world. Gladia offers you the first ten hours of transcriptions for free, after which it costs 0.00017 euros per second.

French pride

A French app comes with a bit of French chauvinism. Gladia is based in Paris and promises ‘100 percent data residency in Europe’ on its website. The French cloud provider OVHCloud announced its intention to acquire Gladia in June, suggesting that both parties have found each other on a technological level as well.

In terms of AI, Gladia does not make the chauvinistic choice. The company has developed its own speech-to-text models, Solaria-1 and Solaria-3. These are built on Whisper, an open-source audio model from OpenAI. That is the only, albeit minimal, non-European trait.

Accurate transcriptions

Gladia allows for various use cases. For example, you can add it to video meetings via the API for transcriptions and translations. This can be done both asynchronously and in real-time. Correctly installing an API does require some technical background. For this test, we are relying on the ‘freemium’ offering of the web version. You get up to ten hours of free transcriptions, but even after that, Gladia is transparent about the costs.

You upload a transcript in one of the supported file formats. The lack of MP4 support is unfortunate, as Microsoft Teams recordings are MP4 by default. Gladia translates the recording sentence by sentence, so you will have to be patient until a complete transcript is ready.

The first recording we test with is an English-language interview. English is the ‘easiest’ language to train AI models on, although in this case, it is not the native language of either speaker, which can influence word pronunciation or sentence structures.

The transcript Gladia produces is a faithful and correct representation of the conversation. We don’t see any gibberish or incomplete sentences anywhere, and every word was attributed to the correct speaker. Only a company name gives every possible spelling except the correct one. You can export your transcripts in written text or other supported file formats (JSON, SRT, VTT).

Except in Dutch

Gladia supports more than a hundred languages and is optimized for the most spoken languages in Europe (English, French, German, Spanish, Italian). In a second test, we run a Dutch-language recording through the tool. We notice almost immediately that this is not the ‘native language’ of the Solaria model.

The transcription process takes a very long time and is full of errors, especially at the beginning. Gladia does recover as more sentences are translated, but the difference in quality between a ‘popular’ language (English) and a less popular language (Dutch) is significant. We cannot conclude whether this is also the case for other languages, but for companies with a local footprint, this is a limitation.

Smart features

Gladia hasn’t put so much effort into developing its own AI models just to have them do parrot work. You can also use the tool for translations of transcriptions or other smart features. We tried out the ‘sentiment analysis’.

Conveying sentiment in a text is not easy, as it can reside in non-verbal cues or subtle hesitations. The ‘umms’ are included in the transcript, but Gladia also tries to represent the atmosphere of the conversation in a graph. Without context, this doesn’t say much; we assume if the graph is in the green section, it was a friendly and amicable conversation. The sentiment analysis in Dutch really goes in all directions, which again shows that reliability is not the same for all languages.

Verdict: more than a transcription tool

Gladia is a strong, versatile transcription tool. Actually, you cannot reduce Gladia to just that, because it supports broader speech-to-text applications. The generous free offer means you can test it out for yourself without being tied down immediately.

If English or French is the primary language in your organization, Gladia will guarantee high-quality transcripts. The tool does not give all languages the same amount of love. If you primarily run Dutch-language interviews or customer conversations through this tool, you will have to be more critical of the output.

ITdaily tests European and open-source alternatives to popular American software tools we use daily. Want to discover more European software? Through this overview you can find all reviews.