In short
A journalist judges a transcription tool on its accuracy and on how it handles recordings. An interview can identify a source. You need to know where the recording is stored, whether it is used to train models and how quotes are linked to the audio. The answers vary from one company to another, and sometimes from one plan to another.
Transcription is one of the tasks where AI saves the most time. Several tools transcribe French and many other languages, and tell speakers apart. Their hosting and data-use terms differ.
What matters to a journalist
Six criteria matter for a journalistic interview:
- transcription quality in the language of the interview, proper names included
- separation of the speakers’ voices
- where recordings are stored and processed
- whether recordings are used to train models
- how long recordings are kept, and how they are deleted
- the step from the transcript to an exact quote
What the companies say
Information collected on 1 October 2026 from the companies’ official pages:
- Otter: hosted on AWS in the United States. Otter trains its own models on de-identified recordings and transcripts. Enterprise plans are excluded by default. Six languages, including French.
- Trint: storage in the European Union or the United States, depending on the account. The Business plan includes a dedicated server in the European Union. Trint says it does not use customer data to train models. More than 40 languages according to its help pages, including French.
- Sonix: its privacy policy states that data is stored outside the European Economic Area and the United Kingdom. Sonix says it does not train models on customer data. It transcribes French.
- Happy Scribe: data centre in the European Union. Training on content requires explicit consent and is off by default. More than 150 languages according to its help pages, including French.
- Descript: information processed and stored in the United States. According to its Security page, sharing data for training is off by default. 26 languages, including French.
Plans change. Links to these pages are listed at the end of the article.
From transcript to quote
An automatic transcript needs checking. A misheard word can change the meaning of a sentence. Before quoting, the journalist listens to the passage again. A published quote reproduces the person’s exact words.
Interviews in Verbatim
Verbatim transcribes MP3, WAV, M4A, AAC, OGG, FLAC and WebM files, as well as videos. It also transcribes recordings made in the app and live interviews, in most languages. It separates the voices and timestamps every passage. No name replaces a speaker until the journalist confirms it.
During a live interview, Verbatim checks every sentence as it is spoken. It compares it with the project’s documents, the sources, the web and everything the person has said before. When an answer contradicts the documents, the excerpt that proves it appears while the person is still speaking. Ahead of the interview, the Prep agent gathers what the person has already said, what the documents show and the questions they will try to avoid.
A quote is copied word for word from the transcript. It is tied to its excerpt only if the text appears there word for word. Each quote has a status: ON by default, or OFF, to quote without naming the source.
Recordings are hosted and processed in Europe. They are never used to train third-party AI.
Questions
What is the best transcription tool for a journalist?
It depends on the interview. For a conversation with no confidentiality issue, a general-purpose tool is enough. For an interview that could identify a source, where the data is stored and whether it is used for training become decisive.
Is automatic transcription reliable?
It can get a proper name, a figure or an indistinct passage wrong. Check every quote against the recording before publication.
Can you record an interview without telling the person?
In France, recording words spoken in a private or confidential capacity without the speaker’s consent is an offence under Article 226-1 of the Criminal Code. Announcing the recording at the start of the interview removes this risk.