Voxtral Small Transcribe 3: Capabilities, Limits, and Cost Analysis

What the specific model does

Voxtral Small Transcribe 3 is a specific variant in Mistral AI's audio catalog, identified as voxtral-small-transcribe-3-0. Unlike the open-weights Voxtral Small 24B model, which is multimodal (audio and text) and supports question-answering, this variant is optimized exclusively for audio-to-text transcription. Its primary goal is to provide high-precision transcription with multilingual support, speaker diarization, and the ability to incorporate specialized vocabulary through contextual biasing.

The model is positioned as a production tool for teams that need to reliably convert audio to text without the computational overhead of a full language model. While Voxtral Small 24B can answer questions about audio content, Voxtral Small Transcribe 3 focuses on transcription fidelity, making it ideal for applications such as subtitling, meeting transcription, or phone call processing.

The distinction between these two variants is about the task each one is presented to perform, not merely about where it runs. Voxtral Small Transcribe 3 takes audio as input and returns a transcript, with diarization and contextual biasing among its stated transcription features. Voxtral Small 24B, by contrast, is described as a multimodal model that can work with audio and text and answer questions about audio. A transcription workflow therefore should not assume that the dedicated transcription variant also provides the broader audio-comprehension functions attributed to the 24B model.

The production positioning describes the intended role of the transcription service, but it does not guarantee a particular error rate or remove the need to assess output. Transcription fidelity can matter differently depending on whether a team is preparing subtitles, records of meetings, or text from phone calls. Those uses share the audio-to-text task, while their requirements for speaker labels, review, and timing may differ. The published feature description supports treating the model as a transcription-focused option; it does not establish that every transcript will be accurate enough for every workflow without checking.

Capabilities and examples by task

The primary capability of Voxtral Small Transcribe 3 is automatic audio transcription in multiple languages. According to Mistral's documentation, the model supports 50 languages for offline transcription, including English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian. The model automatically detects the source language and transcribes the text accordingly, although it also allows forcing a specific language if known in advance.

In addition to basic transcription, the model offers diarization, which identifies and labels different speakers in a conversation. This is crucial for transcribing meetings or interviews where multiple people participate. It also supports contextual biasing, allowing users to provide a list of specialized terms (proper nouns, technical jargon) to improve accuracy in those specific cases.

Automatic language detection and forced-language selection are two ways to specify the language used for transcription. When the language is not supplied, the model is described as detecting it automatically; when it is known in advance, a user can instead force a particular language. These modes are distinct from the published benchmark distinction between inferred and forced languages, so results for one mode should not be assumed to describe the other. The stated support for 50 languages concerns offline transcription, while the listed examples illustrate only some of those languages.

Diarization adds speaker labels to the transcript; it does not, by itself, establish who a speaker is outside the recording or verify the content of what that person says. Its practical value is in separating contributions when more than one person speaks. Contextual biasing addresses a different part of the task: users can supply specialized terms such as names or technical vocabulary to guide transcription in those cases. It may help with the supplied vocabulary, but the feature description does not promise that every supplied term will be recognized correctly or that unrelated transcription errors will disappear.

A contextual bias list is most useful to interpret as guidance about the terms a user expects, rather than as a correction list that guarantees the final spelling. In the software-meeting example, names of APIs, acronyms, and project names are candidates for that guidance. The transcript still needs review where exact terminology matters, because prioritizing terms does not make the transcription deterministic or establish that the audio itself was unambiguous. The example illustrates how a team might apply the feature; it is not a measured accuracy result.

Speaker labels can make an interview transcript easier to read and edit, but labels such as “Interviewer” and “Expert” are role names assigned in the example, not identities the model can establish from the audio alone. Diarization separates speech by speaker; any mapping from those labels to real participants depends on the workflow around the transcript. This distinction also explains why speaker labeling may reduce manual separation work without proving that no additional review or tooling will be needed.

Published benchmarks and methodology

Mistral evaluates the transcription of Voxtral Small Transcribe 3 using an intersection of 45 languages from the FLEURS benchmark. The primary metric is the word error rate (WER) or character error rate (CER), depending on the language's writing system. The provider separates results for forced languages (when the user specifies the language) from inferred languages (when the model detects the language automatically).

For diarization, the cpWER metric (character error rate with speaker penalty) is published across several corpora. It is important to note that these results are provided by the provider and have not been independently replicated by third parties. Users should validate performance with their own data, especially if working with regional accents, background noise, or rapid speaker changes.

WER and CER measure different kinds of transcription differences. WER counts word-level errors, while CER counts character-level errors; the published evaluation uses one or the other depending on the writing system. A score therefore needs to be read alongside its metric and language rather than treated as a single language-independent measure. The separation between forced and inferred language results also matters: it distinguishes runs in which the language is supplied from runs in which the model selects it automatically.

FLEURS supplies the language benchmark named for the transcription evaluation, with the reported intersection covering 45 languages. That evaluation scope is narrower than the stated total of 50 supported offline-transcription languages, so the benchmark description should not be read as a result for all 50. Likewise, benchmark performance describes the evaluated material and setup; it does not by itself predict the result on a team’s own recordings, vocabulary, accents, or recording conditions.

The diarization metric, cpWER, incorporates a speaker penalty, so it reflects transcription errors together with errors associated with speaker assignment. That makes it relevant to evaluating diarized output, but it is not interchangeable with a plain WER or CER score. The published results cover several corpora, and interpreting them requires attention to the metric and corpus rather than a broad claim that speaker labeling will be equally reliable in every conversation. The existing caveat about independent replication and testing on users’ own audio remains important, particularly for regional accents, background noise, and rapid speaker changes.

Pricing and deployment requirements

The price of Voxtral Small Transcribe 3 is $0.004 USD per minute of audio. This rate is public and applies to Mistral's API. There are no charges for text output, as it is included in the audio rate. For private enterprise deployments, a specific agreement with Mistral is required, as no downloadable weights have been published for this specific variant.

Regarding deployment requirements, as an API service, users do not need to manage hardware. However, if considering the base Voxtral Small 24B model for local deployment, approximately 55 GB of GPU RAM is required in bf16 or fp16 format. For Voxtral Small Transcribe 3, hardware requirements are not directly applicable, as it is a managed service.

The stated API rate is based on audio duration, at $0.004 per minute, and the text output is included rather than billed as a separate output charge. This makes the unit of the listed price clear, but a per-minute rate alone does not determine the total cost of a particular project: the amount of audio submitted is also relevant. The public rate applies to the Mistral API as described here; it should not be treated as a hardware price or as the terms of a private enterprise deployment.

The deployment comparison separates a managed transcription API from a locally run open-weights model. Voxtral Small Transcribe 3 is the managed service in this comparison, so its users do not need to provision the model’s inference hardware themselves. Voxtral Small 24B is the local option described as requiring approximately 55 GB of GPU RAM in bf16 or fp16. That estimate belongs to the base 24B model and should not be transferred to the transcription API or interpreted as a requirement for the specific Transcribe 3 variant.

No downloadable weights are published for Voxtral Small Transcribe 3 in the described offering, so the local-deployment comparison concerns Voxtral Small 24B instead. A private enterprise deployment of the transcription variant requires a specific agreement with Mistral; the stated public API price does not establish the terms or availability of such an arrangement. The choice between these deployment paths therefore involves not just the listed rate and hardware estimate, but also whether a managed API or a locally run model fits the intended setup.

Limitations, security, and alternatives

A significant limitation is that the available timestamps are segment-level, not word-level. This may be insufficient for applications requiring precise subtitle synchronization. Additionally, at launch, general audio chat still uses Voxtral 2, so the ability to answer complex questions about audio content should not be attributed to Voxtral Small Transcribe 3.

In terms of security, users must review Mistral's consent and privacy policies, especially if the audio contains personal data. For alternatives, Whisper large-v3 is a free open-source option, but with lower accuracy in some languages. GPT-4o mini Transcribe and Gemini 2.5 Flash are closed alternatives with different prices and capabilities.

Segment-level timestamps associate timing with stretches of transcript rather than with each individual word. That distinction can affect workflows that align text precisely with speech, including subtitle timing. A transcript may still be useful when word-level timing is unnecessary, but the published timestamp granularity does not supply that finer alignment. The limitation concerns timing detail; it should not be confused with the model’s transcription or diarization capabilities.

The note about general audio chat distinguishes the transcription variant from a separate audio-comprehension capability. Voxtral Small Transcribe 3 is described as transcribing audio, while Voxtral 2 is identified as the model still used for general audio chat at launch. A user who needs answers to complex questions about audio should not infer that the transcription service will provide them merely because both products accept audio-related inputs.

Privacy and consent review remain relevant whenever recordings contain personal data. The fact that a service is managed by API does not, on its own, establish that a particular recording is suitable to submit; users need to consider the applicable policies for their use. Similarly, alternatives should be compared on their stated differences in price and capability rather than treated as interchangeable. The characterization of Whisper large-v3 as free and open-source, and of GPT-4o mini Transcribe and Gemini 2.5 Flash as closed alternatives, does not establish that any one option will perform better on a given recording.

The warning about word-level synchronization is a practical consequence of segment-level timestamps, not a claim that the transcript cannot be used for subtitles at all. Where precise alignment is required, the relevant question is whether segment timing meets that requirement. If it does not, a workflow may need another option or additional handling; the limitation alone does not identify which alternative will meet a particular accuracy or timing target.

Practical selection criteria

To decide whether to use Voxtral Small Transcribe 3, consider the following criteria: 1) Do you need pure transcription or also audio comprehension? If you only need transcription, this variant is more efficient. 2) Do you work with technical jargon or proper nouns? If so, contextual biasing is a key advantage. 3) Do you need diarization? If yes, this variant offers it natively. 4) What is your budget? At $0.004 USD/min, it is competitive against other APIs.

If you need audio comprehension capabilities (question-answering, summarization), consider Voxtral Small 24B or alternatives like GPT-4o mini. If cost is the primary priority and accuracy is sufficient, Whisper large-v3 is a free option. For local deployments with strict privacy requirements, Voxtral Small 24B is the open-weights option, although it requires more hardware resources.

Selection begins by separating transcription from audio comprehension. A need to convert speech into text points to the dedicated transcription function; a need to ask questions about audio or produce summaries calls for a model described as supporting those capabilities. The distinction helps prevent a feature mismatch, but it does not establish that one model will be more accurate for every recording or less costly in every deployment. The decision still depends on the task and the way the service will be used.

Diarization and contextual biasing address separate requirements. Speaker labels help organize a conversation by who is speaking, while contextual biasing gives the model vocabulary that may otherwise be difficult to transcribe. A project may need one, both, or neither. Their presence does not guarantee correct speaker assignments or exact recognition of specialized terms, so teams should judge the output against the needs of their own material rather than assuming the feature names alone settle suitability.

The listed $0.004-per-minute API price can be compared with other API prices only when the products’ capabilities and billing terms are also understood. The comparison in the article does not provide a universal cost ranking, and the claim that the price is competitive should not replace a project-specific estimate. For a local deployment, hardware requirements and the need for comprehension features form a different comparison from an API’s per-minute charge.

The final decision process brings together task, features, costs, accuracy checks, and privacy review. Testing with a team’s own audio is especially relevant because the benchmark description does not establish performance for every accent, noise condition, or speaker pattern. Reviewing privacy and security policies is a separate decision from measuring transcription quality: satisfying one criterion does not prove the other. Together, these checks help determine whether the managed transcription service or one of the described alternatives fits the intended workflow without treating any metric, feature, or price as a guarantee.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.