Ask a language model for a dialogue at your level, about the situation you will actually face, then have a text-to-speech model read it aloud in two voices. Replay it, slow it down, check it against a transcript, and ask for a harder version when this one gets easy.
Most listening practice comes as a recording someone else made, at a speed and level someone else picked, about a topic you would never choose. With AI you write the scene instead. A language model drafts a dialogue at the level you ask for, a text-to-speech model reads it aloud in two or more voices, and you press play. It is not a replacement for native speakers, but it fills the long gap between lessons.
Course audio has a ceiling. The dialogues are fixed, the topics are generic, and by the third replay you have memorized the answers instead of training your ear. What most learners need is a steady supply of fresh, short listening material that matches where they are today: a conversation about renting a flat if you are about to look at one, a phone call with a pharmacist if you travel next month, a work meeting if that is where your second language gets used.
AI makes that material cheap to produce. A language model writes the script in whatever level, register and topic you describe, and you can ask for it again with the vocabulary swapped or the pace of the conversation changed. A text-to-speech model then turns the script into audio, and here is the part that matters for listening: you hear a conversation between different voices, not one narrator reading sentences. Picasso IA puts both kinds of model in one place, so the script and the audio come from the same workspace, and the full catalog shows what else is available when you want to try another voice.
The workflow has three parts, and each one is a different kind of model.
The first is the writer. Any strong chat model can draft a dialogue; claude-sonnet-5 and gpt-5 are both in the catalog and both follow instructions about level, length and register well. You describe the situation, the two speakers, the level and the length, and ask for plain text with the speaker names marked, so the script is ready to paste into the next step.
The second is the voice. play-dialog is built for exactly this job: it reads a script as a conversation, supports two voices in one run, and has a language setting. gemini-3.1-flash-tts takes a plain-language style instruction such as speak slowly and clearly, and covers more than seventy language codes, so it suits languages beyond the common ones. elevenlabs v3 offers many voice personas and a language code, which is useful when you want to hear the same line from several different speakers.
The third is optional but worth it: a transcription model such as gpt-4o-transcribe. Feed it the finished audio and compare what it hears with the script. If they differ, either the voice mispronounced something or your ear and the machine both missed a word, and either answer is useful to know.
Level is the single most important thing to put in the writing prompt, and vague words like easy or hard do not work well. Describe the level in terms the model can act on: sentence length, tenses, vocabulary type and how fast people talk. This table is a starting point you can adapt, using the common A1 to C1 labels.
| Level | Ask the language model for | Ask the voice for |
|---|---|---|
| Beginner (A1) | 6 to 8 short lines, present tense only, the most common words, one simple situation such as ordering coffee | Slow pace, clear pauses between turns, one male and one female voice so the speakers are easy to tell apart |
| Elementary (A2) | 8 to 10 lines, past and future in simple forms, everyday topics like shopping or directions | Slightly slower than normal, neutral accent |
| Intermediate (B1) | A 12 to 16 line conversation with opinions, reasons and one small misunderstanding to resolve | Normal pace, a casual style prompt, two clearly different voices |
| Upper intermediate (B2) | A longer exchange with idioms, interruptions and a topic like a work disagreement or a travel problem | Natural pace, expressive delivery, a different accent for each speaker if the model supports it |
| Advanced (C1) | A debate or a three-person scene with implied meaning, humor and fast turn-taking | Fast, informal delivery with minimal pauses |
Two adjustments are worth knowing. First, if the dialogue feels too easy, do not ask for harder words; ask for faster turns, more contractions and more things left unsaid, because real listening difficulty comes from speed and implication more than from rare vocabulary. Second, if it feels too hard, keep the topic and cut the length. A short dialogue you understand teaches more than a long one you only survive.
This walkthrough assumes you have a target language and a rough level. It takes no audio editing and no recording.
Generating the audio is the easy part. What you do with it decides whether your ear improves, and a few habits do most of the work.
Listen in layers. The first pass is only for the situation: who, where, what is going on. The second is for the details you missed. Only the third pass should have the script open. Learners who read along from the start are practicing reading with a soundtrack, which is a different skill.
Change one variable at a time. Keep the same script and swap the voice, or keep the voice and swap the topic. When the same words come back in a new accent or at a new speed, your brain separates the vocabulary from the sound, which is the whole goal of listening practice.
The test of a good listening exercise is not whether you understood it on the first play. It is whether the second play surprised you less than the first. If it did not, the dialogue is too far above your level, and you should ask for a shorter or slower version instead of pushing through.
This is a training aid, and native speakers remain the reference for how a language actually sounds. Five limits to keep in mind:
None of this makes the practice pointless. It means the generated dialogue is your gym and a conversation with a person is the match.
No, and it works best when you do not try to use it that way. A generated dialogue gives you unlimited, level-matched material to train on between lessons, but it cannot reproduce overlapping speech, background noise, regional variety or the unpredictability of a real person. Treat native speakers, real podcasts and real conversations as the reference, and use the AI audio to prepare for them and to fill the gaps in between.
For writing, any strong chat model in the toolkit works; claude-sonnet-5 and gpt-5 are good first choices because they follow level and length instructions closely. For the audio, play-dialog reads a script as a two-voice conversation, gemini-3.1-flash-tts adds a style instruction and many language codes, and elevenlabs v3 offers a wide range of voices. Try the same script in two of them and keep the one your ear finds easiest to follow.
Yes. play-dialog supports two voices in a single run, so you label the lines as Voice 1 and Voice 2 and it reads them as a back-and-forth exchange. For a third speaker, or for a model that handles one voice per run, generate each speaker separately and play the clips in order. Picking voices that differ in pitch or gender makes it much easier to follow who is speaking, especially at beginner level.
That depends on the voice model you choose, and the lists differ. gemini-3.1-flash-tts lists more than seventy language codes, play-dialog has a language setting that covers dozens of languages, and elevenlabs v3 takes a language code too. Check the model page before you commit to a less common language, run a short test, and be ready to switch models if the pronunciation sounds off.
Describe the level in concrete terms rather than with the word easy. Give the CEFR label if you know it, ask for a number of lines, name the tenses you want, and say whether idioms are allowed. If the first version is too hard, ask the model to rewrite it with shorter sentences and more common words; if it is too easy, ask for faster turns and more contractions rather than rare vocabulary.
Yes. Run the finished audio through a transcription model such as gpt-4o-transcribe and compare the output with the script. A difference can mean the voice mispronounced a word or that the audio is unclear, and both are worth knowing. It is also a good self-test: write down what you heard first, then compare your version with the script and the transcript to see exactly which words you missed.
New accounts get free credits, which is enough to try a few dialogues and hear whether the approach suits you. After that, generation costs credits, and the amount depends on the model and the settings. A short text script and a short audio clip are small jobs compared with video, but the exact numbers change, so the pricing page is the place to check before you plan a daily habit around it.
Yes, and it is one of the most useful ways to use it. Paste your list into the writing prompt and ask for a dialogue that uses each word at least once in a natural context, at your level. If a word comes out sounding wrong, ask for another sentence that uses it and listen again.
Short enough to replay several times without losing patience. For most learners that means a handful of lines at beginner level and up to a couple of minutes of audio at advanced level. A one minute dialogue you replay five times does more for your ear than a ten minute one you play once.
For your own learning, a class you teach or a group of friends, generated dialogue with fictional speakers is a practical way to build shared material. Keep the voices fictional: do not try to imitate a real teacher, a celebrity or any identifiable person without their permission. If you plan to publish the audio, check the current terms for the models and platforms you use. For other audio projects, see AI voiceover for videos and audiobook narration.
Pick one situation you will really face this week and open the Picasso IA toolkit to hear it read back at your level.
Choose your language
