La post-correction s’appuie sur un appel à un modèle open source externe, hébergé chez Novita : le texte brut de l’ASR y est traité pour ajouter la ponctuation et les majuscules, corriger les fautes d’orthographe évidentes, et supprimer les répétitions/bégaiements générés par la reconnaissance vocale. Elle ne reformule pas, ne traduit pas et ne modifie pas le sens.
When to use it
- Transcripts intended for a human reader (meeting notes, subtitles, publishable verbatim).
- Long ASR outputs containing hesitations and sentences without punctuation.
- Text that will later be indexed or searched: punctuation and capitalization improve readability and sentence segmentation.
What post-correction does
- Adds capitalization (sentence starts, proper nouns).
- Adds missing punctuation.
- Corrects obvious spelling mistakes.
- Removes immediate repetitions, stutters, and typical ASR glitches.
- Corrects proper nouns only when certainty is high.
- Keeps diarization: any [SPEAKER_xx] markers present in the text are preserved.
What it does not do
- Does not rephrase, does not add content, does not translate.
- Does not change the meaning of the text.
- Does not re-listen to the audio: it corrects only based on the ASR text.
Supported languages
Optimized for French — the same language as the alpha-audio-v1 transcription model. For other languages, do not set enable_postcorrect=true: the text may be altered or returned unchanged with no benefit.
Activation
Simply add enable_postcorrect=true to your multipart request, on the same routes as the standard transcription. No extra field is added to the response: the corrected text is returned directly in the text key of TranscriptResponse.
curl https://api-endpoints.alphaedge-ai.com/models/alpha-audio-v1/transcript \
-H "X-API-Key: TA_CLE" \
-F "audio=@/chemin/audio.wav" \
-F "enable_postcorrect=true"
Before / after example
Raw ASR text (enable_postcorrect absent or false):
aujourd'hui il y a 2 mondes qui nous entourent d'une part la ville étouffante et polluée et d'autre part la forêt qui est une vraie modèle écologique en effet les problèmes majeurs
Post-corrected text (enable_postcorrect=true):
Aujourd'hui, il y a deux mondes qui nous entourent. D'une part, la ville étouffante et polluée, et d'autre part la forêt qui est un vrai modèle écologique. En effet, les problèmes majeurs
Latency
Roughly 1 to 4 seconds per 1,200 characters of transcript. On audios of several minutes, post-correction runs in parallel and typically adds a few dozen seconds to the total wall-clock time. For a ~1-minute audio, the overhead is on the order of 1 to 3 seconds.
The returned inference_seconds field remains the ASR time (reported by the ASR server). Post-correction time is not exposed in the response.
Billing
During the re-enablement phase, post-correction is included at no extra charge in the audio price (€0.15/h). No billing change is required on the client side.
Combining with diarization
enable_diarization and enable_postcorrect are independent and can be combined. When diarization is active, its [SPEAKER_xx] markers are kept in the post-corrected text.
curl https://api-endpoints.alphaedge-ai.com/models/alpha-audio-v1/transcript \
-H "X-API-Key: TA_CLE" \
-F "audio=@/chemin/audio.wav" \
-F "enable_diarization=true" \
-F "enable_postcorrect=true"
Behavior when unavailable
When enable_postcorrect=true is requested but post-correction is not available on the target instance (service unreachable, quota exhausted, server-side disablement), the API returns the raw ASR text with HTTP 200. No error is surfaced to the client: your integration does not need to handle any special case, the text will simply be uncorrected.
For deployments that require a “post-corrected or explicit failure” guarantee, a strict mode returning 502 on failure can be enabled in server configuration — contact support.
Best practices
- For raw verbatim intended for programmatic processing, keep enable_postcorrect=false (default).
- For a human-readable report, set enable_postcorrect=true and, if there are multiple speakers, enable_diarization=true.
- Do not enable post-correction on transcripts in languages other than French.