Artificial IntelligenceTech

Alfred's voice, episode 1: which transcription engine for a meeting assistant?

Léo GaudezLéo Gaudez•2026-11-17•21 min read
Alfred's voice, episode 1: which transcription engine for a meeting assistant?

Alfred is in our meetings

In our Google Meet calls, there is a participant who says nothing until you call it. Its name is Alfred.

I say "Alfred, when is my next meeting?" and it answers out loud, a few seconds after I finish my sentence. And it does more than answer:

  • It searches all of our Notion. A number, a page, a meeting write-up: it finds it.
  • It prepares a meeting in the background. It says "I'm on it" and works for about a minute. Meanwhile, we keep talking. I ask it the time, it answers in two seconds. When the work is ready, it announces it in a lull in the conversation.
  • It shows a page in the meeting and scrolls it by voice.
  • It takes notes: decisions, tasks, a write-up at the end.

It writes nothing to Notion, the calendar or Gmail without a "yes" in the meeting, and it sends no email: at most a draft.

Diagram: Google Meet, speech segmenting, Soniox transcription, router, Sonnet answer with Notion and Google, Gemini voice, back into the Meet microphone

All of this only holds if two things work: it understands what we say, and it answers without making us wait. If either breaks, people stop calling it, and that's the end. I spent three days on it. It makes two episodes:

  • Episode 1 (this one): the ear. Which transcription engine (speech-to-text) understands meeting French, with its English words and jargon, without making me wait.
  • Episode 2 (coming soon): the mouth. Which synthetic voice (text-to-speech), from the first sound to the accent.

My starting point was a free-ish setup: Groq (Whisper) to listen, Kokoro (an open-source voice that runs on the server's CPU) to speak. It worked, nothing more. I wanted to know what I was leaving on the table.

My choice

Soniox, streaming

10.9% word errors, 0.22 s delay, $1.20 for 10 hours of meetings per month. Groq as the fallback.

Compared

≈ 20 engines

About twenty tested, 12 in the chart, on 17 real meeting excerpts and 14 call phrases.

The trap

× 10

The same Soniox takes 2.9 s per segment and 0.2 s streaming.

What I wanted to measure

Four criteria:

  • the word error rate (WER): the share of wrong, missing or extra words, on a normalized text (numbers spelled out, no punctuation, no "uh");
  • a score on 20 business keywords: the terms I do not want to see mangled;
  • the delay: for a streaming engine, between the end of speech and the final text; for a "file" engine, the processing time;
  • the cost for 10 hours of meetings per month.

For an assistant you call out to, delay matters more than I thought. An instruction that is a bit less accurate but ready right away beats a perfect one ten seconds later.

How I tested

Two corpora:

  • 17 excerpts from real meetings (about 9 minutes), with their hesitations, their "uh" and two people talking over each other;
  • 14 call phrases of the "Alfred, …" kind (28 audio files), said by a synthetic voice.

Every engine gets the same audio. I measure WER, keywords, delay and price.

The limits, right away: differences of a few WER points mean nothing on a corpus this small, and the call phrases are not said by colleagues in a hurry.

Three things that made the test harder

1. Nobody wrote down the right answer

To compute an error rate you need a reference text. I had no human transcript of the 17 excerpts, and I was not going to retype them by ear.

So I had several engine families vote: a word kept by the majority goes into the reference. A model then reviewed the result. That is better than nothing, but if every engine makes the same mistake, the reference makes it too. No human validated the text.

2. Per segment or streaming: same engine, ten times slower

At first, Soniox gave me 2.9 seconds of delay, versus under half a second for Groq. That is why I had left Groq as the default. It cost me an evening.

The problem was how I called it. Per segment, I send each finished sentence and wait for the result. Streaming, I send the audio while the person speaks, with one websocket per speaking turn (silence is billed). The text is then ready as soon as the sentence ends: 0.22 s. Same engine, ten times faster.

3. Engines I could not test

  • Azure Speech: account not created.
  • Rev.ai: signup not finished.
  • Cartesia: connection refused.

Twelve engines versus our Franglais

Word error rate and median delay for twelve transcription engines

EngineModeWord errorsKeywordsMedian delayCost, 10 h/month
Gemini 3.8 Flash (audio)file8.2%99%10.55 s$6.50
Soniox stt-rt-v6streaming10.9%96%0.22 s$1.20
AssemblyAI Universal-3.6 Pro Realtimestreaming11.3%90%0.16 s$4.50
ElevenLabs Scribe v2file11.8%96%4.47 s$2.20
Speechmatics Enhancedfile12.3%86%2.31 s$7.50
OpenAI gpt-4o-mini-transcribeper sentence13.3%90%1.24 s$1.80
Gradium STTstreaming13.7%87%0.49 s$0 (free credits)
ElevenLabs Scribe v2 RTstreaming14.6%91%0.20 s$3.90
Mistral Voxtral Mini Transcribefile15.0%87%0.95 s$1.80
Deepgram Nova-3 FRstreaming15.6%88%0.14 s$4.60
faster-whisper large-v3, on CPUlocal16.0%n/a78.99 s$0
Groq whisper-large-v3-turbo (my starting point)per sentence16.2%90%0.46 s$0.40
Google Meet captions (read from the page)streaming27.5%42%0.62 s$0 (included in Meet)

What I take from it:

  • Gemini is the most accurate (8.2%, 99% of keywords), but it takes over 10 seconds. For an instruction spoken to an assistant, that is too long. Its price also comes from reasoning tokens, and it will double in 2027 ($13).

  • Streaming engines win on latency. Soniox, AssemblyAI and Deepgram return the text 0.14 to 0.22 s after the end of the sentence. "File" or "per sentence" engines add 1 to 10 seconds.

  • Mistral Voxtral Mini Transcribe (model voxtral-mini-latest, Mistral's transcription API) is fast on a file (0.95 s) and cheap ($0.003 per minute), but less accurate than the best: 15.0% errors, on par with Deepgram and Groq. It mostly misses proper nouns: "Tailortune" 0 times out of 4, "Bitwarden" 1 out of 3. On the call phrases, it makes 4.2% errors and recognizes "Alfred" every time. I gave it the 20 business terms through context_bias, the way the other engines get their list. The option requires underscores ("Tailor_Studio") that the model copies into its text, so I replace them with spaces before counting. I did not test streaming.

  • Whisper through Groq, my starting point, is the cheapest ($0.40) and among the least accurate: 16.2%.

  • gpt-4o-transcribe, OpenAI's large model, does much worse than its little sibling: 36.3% errors, because it skips passages. It is not in the table.

  • General-purpose multilingual engines (AssemblyAI multilingual, Gladia) truncate and plateau at 23.4%.

  • Google Meet captions are free and fast: the first word appears 1.15 s after the sentence starts, and the final text about 0.5 to 0.6 s after it ends. But at 27.5% errors, Meet ranks second to last of all the engines tested, and last on keywords: Tailor Studio 0 of 6, Tailortune 0 of 4, Bitwarden 0 of 3, and "goals" becomes "golds" or "Gaules". On call phrases, Meet recognizes "Alfred" 15 times out of 16, but makes 18.5% errors. Meet has no captions API and takes no custom vocabulary. So it does not replace Soniox for Alfred, which must recognize its own name and the team's names.

    Google Meet caption on a call phrase read by a synthetic voice: "les rituels c'est le Weekly le point mensuel et le bilan trimestriel des golds" (the rituals are the Weekly, the monthly check-in and the quarterly review of the goals)

    For this test, I played each excerpt into the microphone of a real Meet call and recorded the captions shown. The sound goes through Meet's whole chain, as in real life.

  • Gradium is free up to 45,000 credits per month, about 4 hours, so not enough for 10 hours of meetings. And it adds no punctuation.

Real meeting excerpts

These numbers come from 17 excerpts of our meetings. Here are five, 12 to 18 seconds each, to hear what the engines make of them. I picked them for variety: me alone, a second speaker alone, a conversation between two people, a passage full of English and one where the engines disagree. I left out anything about contracts, clients or personal data.

For each engine, I kept the text up to the same word as the audio: they had transcribed the full original excerpt (18 to 51 seconds). Errors are in red. The red dots mark a forgotten word. I do not mark "uh", filler words ("genre", "ouais") or the spelling of a name like Tailor Studio. The reference comes from the vote described above: no human reviewed it.

Under each engine: the cost of the excerpt (public hourly price × length) and its median delay over the whole corpus.

1. Me alone · 15 s

Me"Bon, en vrai, je sais pas, là j'ai quand même fait un copier-coller du dernier weekly. Mais bon, moi, je continue sur la partie Tailor Studio, migration Notion, migration mail et workspace."

Well, honestly, I don't know, I did copy-paste the last weekly. But anyway, I'm continuing on the Tailor Studio part, Notion migration, mail migration and workspace.

  • Soniox stt-rt-v6 (kept)$0.12/h × 15 s ≈ $0.0005 · delay 0.22 sEuh, bon, en vrai, je sais pas. Là, j'ai quand même fait un copier-coller du dernier weekly. Allez. Euh, mais bon, moi je continue sur la partie Tailor Studio, migration Notion, migration mail et workspace.
  • AssemblyAI Universal-3.6 Pro RT$0.45/h × 15 s ≈ $0.0019 · delay 0.16 sBon, en vrai, je sais pas, là j'ai quand même fait un copier-coller du dernier weekly. Mais bon, moi je continue sur la partie Taylor Studio, migration Notion, migration mail et workspace.
  • ElevenLabs Scribe v2 RT$0.39/h × 15 s ≈ $0.0016 · delay 0.20 sEuh, bon, en vrai, euh, j'sais pas. Là, j'ai quand même fait un copier-coller du dernier weekly. Allez. Euh, mais bon, moi, je continue sur, euh, la partie Taylor Studio, euh, migration Notion, migration mail et workspace.
  • Mistral Voxtral Mini Transcribe$0.18/h × 15 s ≈ $0.0008 · delay 0.95 sBon, en vrai, je ne sais pas. Là, j'ai quand même fait un copier-coller du dernier weekly. Oui. Mais … moi, je continue sur la partie Tailor Studio, migration, notion, migration, mail et workspace.
  • Deepgram Nova-3 FR$0.46/h × 15 s ≈ $0.0019 · delay 0.14 sBon, … vrai, je ne pas, là, j'ai quand même fait un copier-coller du dernier weekly, Mais donc, moi, je continue sur la partie Tailor Studio migration, Notion migration, mail et workspace.
  • Groq whisper-large-v3-turbo$0.04/h × 15 s ≈ $0.0002 · delay 0.46 sBon en vrai, … j'ai quand même fait un copier-coller du dernier weekly. Oui. Mais … moi je continue sur la partie Tailor Studio, migration, Notion, migration, mail et workspace.
  • Gladia$0.73/h × 15 s ≈ $0.0030 · delay 0.24 sbon en vrai je sais pas agent même fait un copier coller du dernier week-end Salut ! Je continue sur la partie... Tailor Studio, Migration, Notion, Migration, Mail et Workspace.

2. Speaker 2 alone · 12 s

Speaker 2"Pour éviter de passer deux heures à discuter en weekly de tous les sujets et que ça dure trop longtemps."

To avoid spending two hours on every topic in the weekly and having it run too long.

  • Soniox stt-rt-v6 (kept)$0.12/h × 12 s ≈ $0.0004 · delay 0.22 sPour éviter de, genre, euh, ouais, passer 2 heures à discuter en weekly de tous les sujets et, euh, et que ça dure trop longtemps.
  • AssemblyAI Universal-3.6 Pro RT$0.45/h × 12 s ≈ $0.0015 · delay 0.16 sPour éviter de passer 2 heures à discuter en weekly de tous les sujets. Et que ça dure … longtemps.
  • ElevenLabs Scribe v2 RT$0.39/h × 12 s ≈ $0.0013 · delay 0.20 sPour éviter de... genre, euh, ouais, passer, euh, deux heures à discuter en weekly, euh, de tous les sujets et, euh... Et que ça dure, euh... …
  • Mistral Voxtral Mini Transcribe$0.18/h × 12 s ≈ $0.0006 · delay 0.95 sPour éviter de passer deux heures à discuter en weekly de tous les sujets. Et que ça dure … longtemps.
  • Deepgram Nova-3 FR$0.46/h × 12 s ≈ $0.0015 · delay 0.14 sPour éviter de genre, ouais, passer 2 heures à discuter en weekly … tous les sujets et et que ça dure …
  • Groq whisper-large-v3-turbo$0.04/h × 12 s ≈ $0.0001 · delay 0.46 spour éviter de passer deux heures à discuter en weekly de tous les sujets et que ça dure … longtemps.
  • Gladia$0.73/h × 12 s ≈ $0.0024 · delay 0.24 spour éviter de... Euh... Genre, ouais, passer deux heures à discuter en weekly de tous les sujets. Oui. Merci. et que ça dure …

3. A conversation between two people · 15 s

Me"Vas-y, on peut passer au goal. Parce que là, franchement, dans l'artifact, c'est bien structuré, tu vois, c'est super visuel. Dans le Notion, c'est dégueulasse."

Speaker 2"Ouais, ouais, non, je suis d'accord avec toi."

Me"Donc, donc, à voir."

Go ahead, we can move on to the goal. Honestly, the artifact is well structured, you see, it's very visual. The Notion one is disgusting. / Yeah, yeah, no, I agree with you. / So, so, we'll see.

  • Soniox stt-rt-v6 (kept)$0.12/h × 15 s ≈ $0.0005 · delay 0.22 sEuh, bon, vas-y, on peut passer aux goals. Parce que là, franchement, … l'artifact, c'est bien structuré, tu vois. C'est super visuel, dans … Notion c'était que là. Ouais, ouais, non, je suis d'accord avec toi. Donc, donc, … voir.
  • AssemblyAI Universal-3.6 Pro RT$0.45/h × 15 s ≈ $0.0019 · delay 0.16 sBon, vas-y, on peut passer au goal. Parce que là, franchement, dans l'artifact, c'est bien structuré, tu vois, c'est super visuel. Dans le Notion, c'était classe. Ouais, ouais, non, je suis d'accord avec toi. Donc à voir.
  • ElevenLabs Scribe v2 RT$0.39/h × 15 s ≈ $0.0016 · delay 0.20 sEuh, bah vas-y, on peut passer au goal. Parce que là, franchement, le rattrapage, c'est bien, euh, structuré, tu vois, euh, c'est super visuel. Dans le Notion, c'est dégueulasse. Ouais, ouais, non, j'suis d'accord avec toi. Donc, euh, donc, y a à voir.
  • Mistral Voxtral Mini Transcribe$0.18/h × 15 s ≈ $0.0008 · delay 0.95 sVas-y, on peut passer au goal. Parce que là, franchement, dans l'artefact, c'est bien structuré, tu vois. C'est super visuel. Dans le motion, c'est dégueulasse. Ouais, ouais, non, je suis d'accord avec toi. Donc, il va y avoir.
  • Deepgram Nova-3 FR$0.46/h × 15 s ≈ $0.0019 · delay 0.14 sBon, vas-y, on peut passer au goals. Parce que là, franchement, on a l'artefact c'est bien structuré, tu vois, c'est super visuel. Dans … Notion, c'est dégueulasse. Ouais ouais, non, je suis d'accord avec toi. Donc, donc, il va voir.
  • Groq whisper-large-v3-turbo$0.04/h × 15 s ≈ $0.0002 · delay 0.46 sBon vas-y on peut passer au goal Parce que là franchement dans l'artifact c'est bien structuré tu vois c'est super visuel dans le notion c'est déculasse Ouais ouais … je suis d'accord avec toi Donc il y a à voir
  • Gladia$0.73/h × 15 s ≈ $0.0030 · delay 0.24 sBon, vas-y, on peut passer au goals. Parce que là, franchement dans … artifact c'est bien structuré, tu vois, c'est super rituels. Dans le Notion c'est dégueulasse. Ouais, ouais, non, je suis d'accord avec toi. Donc, il y a un aurore.

4. Lots of English, me · 18 s

Me"Mais en vrai, moi mon goal, c'est, c'est le workspace ready for Tailor Studio, new launch, content launch."

But honestly, my goal is the workspace ready for Tailor Studio, new launch, content launch.

  • Soniox stt-rt-v6 (kept)$0.12/h × 18 s ≈ $0.0006 · delay 0.22 sDe manière vraie, moi, mon goal, c'est le workspace ready for Tailor Studio new launch. Content launch.
  • AssemblyAI Universal-3.6 Pro RT$0.45/h × 18 s ≈ $0.0023 · delay 0.16 sMais il y a un vrai, moi mon goal c'est le workspace Ready for Taylor Studio … launch. Content launch.
  • ElevenLabs Scribe v2 RT$0.39/h × 18 s ≈ $0.0020 · delay 0.20 sMais en vrai, moi, mon goal, c'est, euh... C'est le workspace, euh, ready for, euh, Tailor Studio, euh, new, euh, launch. Euh, content launch.
  • Mistral Voxtral Mini Transcribe$0.18/h × 18 s ≈ $0.0009 · delay 0.95 sDe manière vraie, … mon goal, c'est... C'est le workspace ready for Tailor Studio … launch. Content launch.
  • Deepgram Nova-3 FR$0.46/h × 18 s ≈ $0.0023 · delay 0.14 smais en vrai, c'est, moi mon goal c'est dès le workspace Ready for Tailor Studio, …
  • Groq whisper-large-v3-turbo$0.04/h × 18 s ≈ $0.0002 · delay 0.46 s… En vrai, moi mon goal c'est le workspace ready for Tailor Studio … launch, content launch.
  • Gladia$0.73/h × 18 s ≈ $0.0037 · delay 0.24 sje vais dire vrai. Moi, mon goals, c'est... C'est le workspace. C'est... Merci. Merci. Merci. content l'homme de chaque minute.

5. When the engines disagree · 17 s

Me"Attends, je te laisse parler à l'agent. Faut juste que je réponde à ça. Allô ?"

Speaker 2"Ouais, là, ça a pas mis, ça a pas mis le monthly, pilotage des réunions. Ça a pas mis le monthly, etc."

Hold on, I'll let you talk to the agent. I just have to answer this. Hello? / Yeah, so, it didn't put, it didn't put the monthly, meeting management. It didn't put the monthly, etc.

  • Soniox stt-rt-v6 (kept)$0.12/h × 17 s ≈ $0.0006 · delay 0.22 sPardon. Je te laisse parler à l'agent, je veux juste que je réponde à ça. Allô ? Ouais. Là, ça a pas mis, euh, ça a pas mis le monthly. Pilotage des réunions. Euh, ça a pas mis le monthly, etc.
  • AssemblyAI Universal-3.6 Pro RT$0.45/h × 17 s ≈ $0.0021 · delay 0.16 s… Je te laisse parler à l'agent, je réponds juste à ça. Allô? Ouais, là ça a pas mis le monthly. Puis le test d'avant-nous. Ça a pas mis le mensuel, etc.
  • ElevenLabs Scribe v2 RT$0.39/h × 17 s ≈ $0.0018 · delay 0.20 sOuais, donc, c'tu veux que j'te laisse parler à l'agent ? J'fous juste-- j'réponds à ça. Euh, allô ? Ouais, là, ça a pas mis, euh, ça a pas mis le monthly. Pilotage des réunions. Euh, ça a pas mis le monthly, et cetera.
  • Mistral Voxtral Mini Transcribe$0.18/h × 17 s ≈ $0.0009 · delay 0.95 s… Je te laisse parler à l'agent, je réponds à ça. Allô ? Là, ça n'a pas mis le monthly. Pilotage derrière nous. Ça n'a pas mis le monthly, etc.
  • Deepgram Nova-3 FR$0.46/h × 17 s ≈ $0.0022 · delay 0.14 sAttends, je te laisse parler à l'agent, veux juste, je réponds à ça. Allô Ouais, là ça n'a pas mis, ça n'a pas mis le Monsley, pilotage des réunions. Ça a pas mis le Monsley, et caetera.
  • Groq whisper-large-v3-turbo$0.04/h × 17 s ≈ $0.0002 · delay 0.46 s… Je te laisse parler à l'agent, je peux juste répondre à ça. Allô ? Ouais, là ça n'a pas mis le monthly, pilotage derrière nous. Ça n'a pas mis le monthly, etc.
  • Gladia$0.73/h × 17 s ≈ $0.0034 · delay 0.24 s… Je te laisse parler … agent, je vous raconte ça. Allô ? Là, ça n'a pas mis... Ça n'a pas mis le monthly. pilotage dernier. sa famille, le menselier, etc.

Call phrases

Two phrases of the "Alfred, …" kind, said by a synthetic voice. A "3" in place of "trois" counts as correct: my scores normalize digits.

Phrase A. "Alfred, rappelle-nous les trois goals Q4 et le statut de chaque feature." (Alfred, remind us of the three Q4 goals and the status of each feature.)

  • Soniox stt-rt-v6 (kept)Alfred, rappelle-nous les 3 goals qu'il faut et le statut de chaque feature.
  • AssemblyAI Universal-3.6 Pro RTAlfred rappelle où les trois goals qu'il faut et le statut de chaque. Feature.
  • ElevenLabs Scribe v2 RTAlfred, rappelle-nous les trois goals Q4 et le statut de chaque feature.
  • Deepgram Nova-3 FRAlfred, rappelle-nous les 3 goals QFO et le statut de chaque Fitch.
  • OpenAI gpt-4o-mini-transcribeAlfred, rappelle-nous les trois goals qu'il faut et le statut de chaque feature.
  • Groq whisper-large-v3-turboAlfred, rappelle-nous les trois goals Q4 et le statut de chaque fiche.

Phrase B. "Alfred, projette l'artifact du pricing et scrolle jusqu'au benchmark." (Alfred, show the pricing artifact and scroll to the benchmark.)

  • Soniox stt-rt-v6 (kept)Alfred, projette l'artifact du pricing et scrolle jusqu'au benchmark.
  • AssemblyAI Universal-3.6 Pro RTAlfred projette l'artefact du pricing et scrolle jusqu'au benchmark.
  • ElevenLabs Scribe v2 RTAlfred projette l'artefact du pricing et scrolle jusqu'au benchmark.
  • Deepgram Nova-3 FRAlfred projette l'artefact du pricing et scrolle jusqu'au benchmark.
  • OpenAI gpt-4o-mini-transcribeAlfred projette l'artifact du pricing et scroll jusqu'au benchmark.
  • Groq whisper-large-v3-turboAlfred projette l'artifact du pricing et scrolle jusqu'au benchmark.

Why Soniox

Between Soniox and AssemblyAI, 10.9% versus 11.3% is measurement noise. What breaks the tie:

  • price: $1.20 versus $4.50 for 10 hours, about four times less;
  • keywords: 96% versus 90%, thanks to a context of 20 business terms and the meeting topic given to the engine;
  • delay: 0.22 s on my test bench, about 0.5 s once plugged into the bot, end-of-sentence detection included.

On the call phrases, Soniox makes 2% errors versus 4.4% for Groq, and "Alfred" is recognized nearly every time.

What it feels like with Alfred

Before, transcription alone took 2.9 s. Now the text is ready half a second after the end of my sentence, with a third fewer errors than my starting point, for $1.20 a month. From the end of my question to Alfred's voice, it takes about 5 s, and about 3 s with the next version of the router, already in rehearsal.

Now it has to speak without making anyone wince. That's episode 2.

An article about Alfred?

Alfred deserves more than a detour through its voice. If you're interested, write to me and I'll do a dedicated article on Alfred: how it is built, how it helps us work together and how it navigates our Notion.

Write to me