Research Scientist, Speech & Audio

Innodata Inc., Remote - United States +1

What you will do

You will define how Innodata designs, structures, and evaluates audio data for speech and audio models, and you will validate those choices experimentally. Concretely, you will:
Translate the requirements of speech and audio models — ASR, text-to-speech and speech generation, speech-to-speech and conversational voice, speaker diarization and verification, audio-language models, and streaming systems — into concrete data specifications: modalities, transcription and annotation schemas, sampling, and evaluation criteria.
Build evaluation methodology that goes past word error rate — semantic accuracy, robustness to noise and accent, code-switching, diarization error rate (DER), naturalness and intelligibility of generated speech, and streaming latency — and know when automated metrics hold and when they don't.
Decide how existing and incoming audio should be structured, enriched, and sampled for coverage that fits the model objective — across languages, accents, and acoustic conditions (studio, real-world, telephonic), speaker demographics, emotional and paralinguistic range, scripted versus spontaneous speech, and single- versus multi-speaker settings, including low-resource and code-switched speech.

Innodata Inc., Remote - United States +1

You will define how Innodata designs, structures, and evaluates audio data for speech and audio models, and you will validate those choices experimentally. Concretely, you will:
Translate the requirements of speech and audio models — ASR, text-to-speech and speech generation, speech-to-speech and conversational voice, speaker diarization and verification, audio-language models, and streaming systems — into concrete data specifications: modalities, transcription and annotation schemas, sampling, and evaluation criteria.
Build evaluation methodology that goes past word error rate — semantic accuracy, robustness to noise and accent, code-switching, diarization error rate (DER), naturalness and intelligibility of generated speech, and streaming latency — and know when automated metrics hold and when they don't.
Decide how existing and incoming audio should be structured, enriched, and sampled for coverage that fits the model objective — across languages, accents, and acoustic conditions (studio, real-world, telephonic), speaker demographics, emotional and paralinguistic range, scripted versus spontaneous speech, and single- versus multi-speaker settings, including low-resource and code-switched speech.